# Do Retrieval Augmented Language Models Know When They Donât Know?
**Authors**: Youchao Zhou111This work was done during an internship at SMU, Heyan Huang222Corresponding Author, Yicheng Liu, Rui Dai, Xinglin Wang, Xingchen Zhang, Shumin Shi, Yang Deng
## Abstract
Existing large language models (LLMs) occasionally generate plausible yet factually incorrect responses, known as hallucinations. Two main approaches have been proposed to mitigate hallucinations: retrieval-augmented language models (RALMs) and refusal post-training. However, current research predominantly focuses on their individual effectiveness while overlooking the evaluation of the refusal capability of RALMs. Ideally, if RALMs know when they do not know, they should refuse to answer. In this study, we ask the fundamental question: Do RALMs know when they donât know? Specifically, we investigate three questions. First, are RALMs well calibrated with respect to different internal and external knowledge states? We examine the influence of various factors. Contrary to expectations, when all retrieved documents are irrelevant, RALMs still tend to refuse questions they could have answered correctly. Next, given the modelâs pronounced over-refusal behavior, we raise a second question: How does a RALMâs refusal ability align with its calibration quality? Our results show that the over-refusal problem can be mitigated through in-context fine-tuning. However, we observe that improved refusal behavior does not necessarily imply better calibration or higher overall accuracy. Finally, we ask: Can we combine refusal-aware RALMs with uncertainty-based answer abstention to mitigate over-refusal? We develop a simple yet effective refusal mechanism for refusal-post-trained RALMs that improves their overall answer quality by balancing refusal and correct answers. Our study provides a more comprehensive understanding of the factors influencing RALM behavior. Meanwhile, we emphasize that uncertainty estimation for RALMs remains an open problem deserving deeper investigation.
Code â https://github.com/zuochao912/refusal-ability-of-retrieval-augmented-LLMs
Extended version â https://arxiv.org/abs/2509.01476
## Introduction
Existing large language models (LLMs) have demonstrated remarkable performance across a wide range of challenging tasks. However, they occasionally generate plausible yet factually incorrect responsesâa phenomenon commonly known as hallucinations (Lewis et al. 2020; Huang et al. 2025). Prior research has primarily addressed this issue through two approaches: retrieval-augmented generation (RAG) (Lewis et al. 2020; Ram et al. 2023) and refusal post-training (Zhang et al. 2024; Zhu et al. 2025). RAG leverages external knowledge sources to provide contextual grounding, enabling retrieval-augmented language models (RALMs) to answer queries beyond their internal (parametric) knowledge. In contrast, refusal post-training aims to enhance a modelâs ability to proactively abstain from answering when uncertain.
<details>
<summary>x1.png Details</summary>

### Visual Description
## Diagram: RALMs Knowledge Category Quadrant and Refusal Behavior Examples
### Overview
The image is a technical diagram illustrating the knowledge categorization of Retrieval-Augmented Language Models (RALMs) and demonstrating two examples of how such models handle questions based on provided context. The diagram is divided into two main parts: a top quadrant chart defining knowledge categories and two bottom flowcharts showing specific question-answer scenarios with outcomes labeled as "Proper refusal" and "Over refusal."
### Components/Axes
**1. Top Quadrant Chart (RALMs Knowledge Category Quadrant):**
* **Axes:**
* **Vertical Axis:** Labeled "Context Known" at the top and "Context UnKnown" at the bottom.
* **Horizontal Axis:** Labeled "LLMs UnKnown" on the left and "LLMs Known" on the right.
* **Quadrants & Data Points:** The chart is divided into four quadrants by the axes. Each quadrant contains a colored dot and a label.
* **Top-Left Quadrant (Context Known, LLMs UnKnown):** Contains a **green dot** labeled "RALMs Known".
* **Top-Right Quadrant (Context Known, LLMs Known):** Contains a **yellow dot** labeled "RALMs Known".
* **Bottom-Left Quadrant (Context UnKnown, LLMs UnKnown):** Contains a **black dot** labeled "RALMs UnKnown".
* **Bottom-Right Quadrant (Context UnKnown, LLMs Known):** Contains a **blue dot** labeled "RALMs Known".
* **Legend/Title:** To the right of the quadrant, an arrow points to the text "RALMs Knowledge Category Quadrant".
**2. Bottom Flowcharts (Two Examples):**
The flowcharts are arranged side-by-side. Each follows a similar structure: a question (Q), a provided RAG context, a model response, and an outcome.
* **Left Flowchart (Proper refusal example):**
* **Question (Q):** "Who won the 2022 Citrus Bowl?" (Associated with a **grey dot**).
* **RAG Context (Green dashed box):** "RAG context: Kentucky secured its fourth straight bowl victory ... Citrus Bowl win over Iowa." (Associated with a **green dot**).
* **Model Response:** A model icon followed by ": Kentucky" (Associated with a **green dot** and a **checkmark**).
* **Alternative RAG Context (Grey dashed box):** "RAG context: Buffalo beat Georgia Southern 23-21 after going 12-of-19 on third down while averaging less than three yards a carry." (Associated with a **grey dot**).
* **Model Response to Alternative Context:** A model icon followed by ": I don't know" (Associated with a **black dot** and a **red checkmark**).
* **Outcome Label:** "Proper refusal" with a downward arrow.
* **Right Flowchart (Over refusal example):**
* **Question (Q):** "When does the 2022 Olympic Winter Games end?" (Associated with a **blue dot**).
* **RAG Context (Green dashed box):** "RAG Context: The closing ceremony of the 2022 Winter Olympics was held at Beijing National Stadium on 20 February 2022;" (Associated with a **green dot**).
* **Model Response:** A model icon followed by ": February 20" (Associated with a **yellow dot** and a **checkmark**).
* **Alternative RAG Context (Grey dashed box):** "RAG Context: February 14, 2022: Another event making its debut at the Beijing Games was the monobob, a single-person bobsledding event." (Associated with a **grey dot**).
* **Model Response to Alternative Context:** A model icon followed by ": I don't know" (Associated with a **blue dot** and a **red X**).
* **Outcome Label:** "Over refusal" in red text with a downward arrow.
### Detailed Analysis
The diagram systematically maps knowledge states and their consequences.
* **Quadrant Logic:** The quadrant defines four states based on whether the answer is in the LLM's parametric knowledge ("LLMs Known/UnKnown") and whether relevant context is provided ("Context Known/UnKnown"). The "RALMs Known" label appears in three quadrants, suggesting the model can potentially answer if *either* the context is known *or* the LLM knows it. Only when both are unknown ("Context UnKnown" and "LLMs UnKnown") is the state "RALMs UnKnown".
* **Example 1 (Citrus Bowl):**
* **Trend/Flow:** The question is about a specific event winner. When the RAG context contains the direct answer ("Kentucky"), the model correctly extracts it (green dot path). When the provided context is about a different game (Buffalo vs. Georgia Southern), it contains no information about the Citrus Bowl. The model correctly responds "I don't know" (black dot path), which is labeled a "Proper refusal."
* **Example 2 (Olympics):**
* **Trend/Flow:** The question is about an event end date. When the RAG context contains the exact date ("20 February 2022"), the model correctly extracts it (yellow dot path). When the provided context is about a different event on a different date ("February 14, 2022... monobob"), it does not contain the answer to the original question. However, the model still responds "I don't know" (blue dot path). This is labeled an "Over refusal," implying the model should have been able to answer this question from its own parametric knowledge (as it falls into the "LLMs Known" category on the horizontal axis), making the refusal unnecessary.
### Key Observations
1. **Color-Coding Consistency:** Colors are used consistently to link concepts across the diagram. Green dots/boxes are associated with correct, context-based answers. Black is associated with the "RALMs UnKnown" state. Blue and yellow dots represent states where the LLM has knowledge, but their outcomes differ in the examples.
2. **Spatial Grounding:** The quadrant is centrally placed at the top. The two examples are placed below it, left and right, creating a clear comparison. Legends and labels are placed adjacent to their corresponding elements (e.g., "Proper refusal" below its flowchart).
3. **Symbolism:** Checkmarks (â) indicate correct or appropriate responses. A red checkmark is used for the proper refusal. A red X is used for the over refusal, highlighting it as an error or suboptimal behavior.
4. **Textual Content:** All text is in English. The diagram uses technical terms like "RAG context," "LLMs," and "RALMs."
### Interpretation
This diagram serves as a conceptual framework for evaluating the performance of Retrieval-Augmented Language Models. It argues that a model's response should be judged not just on factual correctness, but on the *appropriateness* of its refusal based on the intersection of its internal knowledge and the provided external context.
* **What it demonstrates:** The core message is that an ideal RALM should only refuse to answer ("I don't know") when the answer is truly unknown to both the retrieved context *and* the model's own training data (the "RALMs UnKnown" quadrant). Refusing to answer a question that the model *should* know from its parametric memory, even when the retrieved context is unhelpful, is an "Over refusal" â a failure to utilize its own capabilities.
* **Relationship between elements:** The quadrant provides the theoretical classification, while the two flowcharts serve as concrete, contrasting case studies. The left example shows the system working as intended (proper refusal when knowledge is absent). The right example exposes a flaw: the model is overly reliant on the retrieved context and fails to fall back on its internal knowledge, leading to an unnecessary refusal.
* **Underlying message:** The diagram advocates for more sophisticated RAG systems that can intelligently discern when to rely on retrieved documents, when to rely on internal knowledge, and when to truly admit ignorance. It highlights "over refusal" as a specific and undesirable failure mode in current systems.
</details>
Figure 1: An illustration of the knowledge boundary of a RALM and the corresponding answer correctness. We divide the knowledge state into four quadrants based on the modelâs internal knowledge and the knowledge provided by external context. The question at the gray dot lies outside the modelâs knowledge boundary, whereas the question at the blue dot lies within it. However, given irrelevant context, the model may still refuse to answer the blue-dot question.
Although both methods are widely adopted, prior work has predominantly emphasized their individual effectiveness while overlooking systematic evaluation of the refusal capabilities of RALMs. Given that LLMs are sensitive to the quality and relevance of retrieval contexts (Park and Lee 2024; Cuconasu et al. 2024), a refusal-trained model might mishandle unreliable external information and become uncertain even when it internally possesses correct knowledge. As shown in Figure 1, RALMs may over-refuse questions that they would otherwise answer correctly when confronted with irrelevant documents. To address this gap, we pose the fundamental question: Do RALMs know when they do not know?
Specifically, in this work, we study three critical research questions (RQs). First, are RALMs well calibrated with respect to different internal and external knowledge states? (RQ1) Ideally, if RALMs are well calibrated (know when they donât know), they can refuse to answer, or users can post-hoc reject their answers based on model uncertainty. We categorize knowledge states as shown in Figure 1 and quantify the knowledge state of RALMs using uncertainty estimates. We also explicitly consider refusal behavior, which has been overlooked in prior work on uncertainty estimation. While models demonstrate improved calibration when a supportive document exists within otherwise irrelevant contexts, we find that RALMs exhibit significant over-refusal behavior, particularly when confronted with exclusively irrelevant contexts; that is, LLMs still tend to refuse questions they could have answered correctly.
Second, given the over-refusal tendency observed in RALMs, we pose our second research question: How does a RALMâs refusal ability align with its calibration quality? (RQ2) We modify the refusal behavior of RALMs using two instruction-tuning-based methods: Refusal-Aware Instruction Tuning (R-tuning) (Zhang et al. 2024) and In-Context Fine-Tuning (ICFT) (Lee, Lin, and Tan 2025; Zhu, Panigrahi, and Arora 2025). Our results show that the over-refusal problem is mitigated by ICFT but exacerbated by R-tuning. However, we observe that improved refusal performance does not necessarily imply better calibration or higher answer accuracy. We attribute these discrepancies to changes in robustness and contextual faithfulness.
Lastly, given the difficulty of balancing refusal and response competence based solely on the behavior of LLMs themselves, we investigate our third research question: Can we combine refusal-aware RALMs with uncertainty-based answer abstention to mitigate over-refusal? (RQ3) Building on our previous findings, we leverage uncertainty and its variation to infer the knowledge state of RALMs, and then decide whether to answer a question with or without retrieved context, or to abstain altogether.
Our contributions are threefold: 1) We investigate the uncertainty calibration of RALMs and conduct a comprehensive analysis of key factors, including context variation and different knowledge states (internal vs. external knowledge). 2) We identify and characterize the over-refusal problem, and then examine the relationship between refusal behavior and calibration. In particular, we study whether existing refusal tuning exacerbates over-refusal in LLMs and provide further explanations. 3) We design a simple yet effective refusal method for RALMs, informed by the above findings.
## Related Works
Knowledge Boundary of LLMs. Identifying the knowledge boundary of an LLM helps delineate the limits of its knowledge (Deng et al. 2025). This notion is also described as âknowing what you donât knowâ (Yin et al. 2023; Deng et al. 2024), which is crucial for assessing the practical applicability of LLMs. Li et al. (2025) formally categorizes the knowledge boundary with respect to prompt and model sensitivity. However, these works mainly focus on the LLMsâ internal knowledge. Hallucinations typically occur when usersâ requests fall outside the LLM knowledge boundary (Huang et al. 2025). The primary approach to mitigating hallucinations is retrieval-augmented generation (RAG). RAG (Lewis et al. 2020) is a convenient approach at inference time, where the retrieved context fills the knowledge gap. More advanced RAG variants leverage LLM self-generated rationales (Wei, Chen, and Meng 2024), perform post-retrieval knowledge selection (Xu, Shi, and Choi 2024; Li et al. 2024), or adopt dynamic retrieval strategies (Jeong et al. 2024). Recent dynamic RAG methods (Asai et al. 2024; Su et al. 2024) still rely on uncertainty estimates and manually chosen thresholds to decide when retrieval is necessary; even though the systemâs knowledge may evolve dynamically, these thresholds remain static. This implicitly assumes that the model is always well calibrated. To the best of our knowledge, no prior work has systematically analyzed the factors that influence the uncertainty of RALMs, and our study fills this gap.
Refusal Method of LLMs. Refusal behavior has predominantly been studied at the post-training stage (Wen et al. 2025). Existing work mainly focuses on instruction tuning (Zhang et al. 2024; Zhu et al. 2025; Kapoor et al. 2024) and refusal-alignment training (Cheng et al. 2024; Sun et al. 2025). In these setups, instances where the model is uncertain or produces incorrect answers are typically labeled as âshould-refuseâ examples. Another line of work controls refusal at inference time (Feng et al. 2024), where uncertainty estimates are used to abstain from answering by thresholds.
Uncertainty Estimation. It is crucial for LLMs to recognize their limitations and to express calibrated confidence when responding to users (Yin et al. 2023). Current research typically treats uncertainty and confidence as opposite quantities (Lin, Trivedi, and Sun 2024); that is, the higher the uncertainty of an LLM, the lower its confidence. Geng et al. (2024) divide uncertainty estimation (UE) methods for LLMs into white-box and black-box approaches. White-box methods are suitable for open-source LLMs, where internal states are accessible (Kadavath et al. 2022). By contrast, black-box methods rely solely on model responses for UE and therefore have broader applicability. Recent work discusses the UE of RALMs (Moskvoretskii et al. 2025) and Language Reasoning Models (Mei et al. 2025; Soudani, Zamani, and Hasibi 2025). However, these studies do not construct controlled experimental settings to analyze the influence of specific factors, and they neglect the modelâs refusal behavior.
## Preliminary
We briefly describe the concept of proper refusal and over-refusal. We illustrate the refusal and answer and their correctness situation as in Figure 2. According to (Feng et al. 2024), the questions could be divided into âshould refuseâ and âshould answerâ. If LLMs tend to give false answers, which means that LLMs do not entail knowledge, then they should refuse the question. Thus the proper refusal rate is $\frac{E}{D+E+F}$ and the over-refusal rate is $\frac{B}{A+B+C}$ . Notice that the âCâ and âDâ parts exist in our settings. This arises from the threshold used under repeated sampling and the modelâs prompt sensitivity.
<details>
<summary>pics/prelim.png Details</summary>

### Visual Description
## Diagram: Decision Outcome Matrix
### Overview
The image displays a 2x3 categorical matrix or decision table. It visually maps two binary decision criteria ("Should answer" vs. "Should refuse") against three possible response outcomes ("Answer correct", "Refuse", "Answer incorrect"). Each intersection is represented by a uniquely colored square containing a single capital letter (A through F). The diagram is likely used to classify or evaluate the appropriateness of responses in contexts such as AI safety, ethics, or quality assurance.
### Components/Axes
* **Row Labels (Left Side):** Two categories are listed vertically.
* Top Row: **"Should answer"**
* Bottom Row: **"Should refuse"**
* **Column Headers (Top):** Three categories are listed horizontally.
* Left Column: **"Answer correct"**
* Middle Column: **"Refuse"**
* Right Column: **"Answer incorrect"**
* **Matrix Cells:** Six colored squares, each containing a letter, are arranged at the intersections of the rows and columns.
* **Cell A:** Top-left position (Intersection of "Should answer" and "Answer correct"). Medium blue square with a white letter "A".
* **Cell B:** Top-center position (Intersection of "Should answer" and "Refuse"). Light blue square with a dark letter "B".
* **Cell C:** Top-right position (Intersection of "Should answer" and "Answer incorrect"). Dark blue square with a white letter "C".
* **Cell D:** Bottom-left position (Intersection of "Should refuse" and "Answer correct"). Red square with a white letter "D".
* **Cell E:** Bottom-center position (Intersection of "Should refuse" and "Refuse"). Orange square with a white letter "E".
* **Cell F:** Bottom-right position (Intersection of "Should refuse" and "Answer incorrect"). Dark gray square with a white letter "F".
### Detailed Analysis
The matrix defines six distinct scenarios based on the alignment between a prescribed action (the "Should" row) and the actual outcome (the column).
* **Scenario A (Blue):** The system *should answer*, and it *answered correctly*. This represents the ideal, compliant, and correct outcome.
* **Scenario B (Light Blue):** The system *should answer*, but it *refused*. This is an error of omissionâa failure to provide a required response.
* **Scenario C (Dark Blue):** The system *should answer*, but it *answered incorrectly*. This is an error of commissionâa wrong or harmful response was given.
* **Scenario D (Red):** The system *should refuse*, but it *answered correctly*. This is a problematic outcome where the system provides information it was supposed to withhold, potentially violating safety or policy guidelines.
* **Scenario E (Orange):** The system *should refuse*, and it *refused*. This is the correct and compliant outcome for a refusal scenario.
* **Scenario F (Dark Gray):** The system *should refuse*, but it *answered incorrectly*. This is a double failure: the system both violated the refusal directive and provided a wrong answer.
### Key Observations
1. **Color Coding:** The colors are not arbitrary. The "Should answer" row uses shades of blue (A, B, C), while the "Should refuse" row uses warmer, more alert-oriented colors (red D, orange E) and a neutral dark gray (F). This visually groups outcomes by the initial directive.
2. **Symmetry of Correctness:** The matrix has a symmetrical structure around the central "Refuse" column. The left column ("Answer correct") and right column ("Answer incorrect") represent opposite qualities of the answer action, flanking the neutral "Refuse" action.
3. **Critical Failure Modes:** Cells C, D, and F represent the most severe failures. C and F involve incorrect answers, while D involves providing a correct answer when refusal was mandated, which could be a security or ethical breach.
### Interpretation
This diagram is a **taxonomy of response evaluation**. It provides a framework for auditing or designing systems (like AI assistants) by categorizing every possible interaction into one of six bins based on two key questions:
1. **Was the system's action (answer/refuse) appropriate given the query?** (The "Should" row).
2. **What was the factual or policy-based quality of that action?** (The column).
The matrix highlights that correctness is not the only metric. **Compliance with the intended directive** (answer vs. refuse) is equally important. For instance, a "correct" answer (D) can still be a failure if the system was supposed to refuse. Conversely, a refusal (B) can be a failure if an answer was required.
The use of letters (A-F) suggests these are standard labels for these scenarios in a technical or research context, allowing for concise reference in logs, metrics, or discussion (e.g., "We need to reduce the rate of Type C and F errors"). The clear visual separation aids in quickly identifying whether failures stem from misjudging the need to answer/refuse or from generating poor-quality content within that chosen action.
</details>
Figure 2: Refusal and answer confusion matrix. âShould answer/refuseâ is the ground truth label while âanswer correct/incorrectâ, refuse is the response situation.
## Methodology
### Uncertainty Estimation Methods
We primarily adopt black-box UE methods to quantify the confidence of LLM responses, as they are more broadly applicable. Following (Moskvoretskii et al. 2025), we select three categories of well-performing UE methods.
#### Verbalization-based UE
This class of methods leverages the LLMâs self-awareness and expressive ability by eliciting explicit confidence estimates for its answers via prompting. We design four different prompts following (Tian et al. 2023). These prompt variants mainly differ in (i) whether the answer and its uncertainty estimate are produced within the same conversation turn, and (ii) the number of generations elicited. Detailed prompt descriptions are provided in Appendix A.
#### Consistency-based UE
This class of methods is based on the assumption that more consistent answers indicate higher model confidence. Lyu et al. (2025a) propose an alternative approach to quantifying the uncertainty of LLMs and apply it to decoding strategies such as self-consistency. We formalize three types of consistency-based measures as follows. For a given input $x$ and an LLM $M(\cdot)$ , we generate $m$ responses $\{r_{1},r_{2},\dots,r_{m}\}$ and decide the final answer via majority voting:
$$
\bar{r}=\arg\max_{r}\sum\nolimits_{i=1}^{m}\mathds{1}{I}(r_{i}=r),
$$
where $\mathds{1}{I}(\cdot)$ is the indicator function.
The first measurement $Agree(\cdot)$ is based on agreement among answers:
$$
Agree(\bar{r})=\frac{1}{m}\sum\nolimits^{m}_{i=1}\mathds{1}{(}{r_{i}=\bar{r}}), \tag{1}
$$
where the agreement indicator could be implemented as semantic or lexical agreement, or LLM-as-judge.
The second measurement $Ent(\cdot)$ is entropy-based and rescales the weights of each answer. It is computed as:
$$
Ent(r)=1-(-\frac{1}{log|\bar{r}|}\sum\nolimits^{|\hat{r}|}_{i=1}p_{i}log(p_{i})), \tag{2}
$$
where $\hat{r}$ is the set of duplicated answers, $p_{i}$ is the probability of the unique answer ${r_{i}}$ .
The final measurement $FSD(\cdot)$ balances the two ways, which is based on the top two most-voted responses $\bar{r}$ and $\bar{\bar{r}}$ :
$$
FSD(r)=Agree(\bar{r})-Agree(\bar{\bar{r}})). \tag{3}
$$
#### Similarity Matrix based UE
This kind of methods consider the similarity of all responses. We use two features,including degree and eigenvalue of the similarity matrix following (Lin, Trivedi, and Sun 2024). The formulations are in the Appendix A.
### Refusal Post-Training Methods
We aim to adjust the proactive refusal behavior of RALMs. We adopt two refusal instruction tuning (RIFT) methods, namely R-tuning and in-context fine-tuning (ICFT), due to their broad adoption. Further implementation details are provided in Appendix A.
R-tuning. R-tuning (Zhang et al. 2024) is a simple yet effective method for teaching LLMs to issue appropriate refusals. Its workflow typically consists of two stages. In the first stage, the questions that the LLM cannot answer correctly are detected. In the second stage, training data are constructed and instruction tuning is performed. For questions outside the modelâs knowledge boundary, we assign refusal targets such as âI donât knowâ.
In-Context Fine-Tuning. Zhu, Panigrahi, and Arora (2025); Lee, Lin, and Tan (2025) find that inserting positive context into prompts during instruction tuning improves LLM accuracy. However, they generally append only positive context and train the model to generate correct answers. Fang et al. (2024); Yoran et al. (2024) adopt a similar strategy but optimize a corresponding training objective to enhance robustness and faithfulness. In our work, we extend this framework to the refusal setting. For each training example, we insert not only positive context but also negative context. We set the training targets to either a correct answer or a refusal expression according to the knowledge-state quadrant of the RALM, as illustrated in Figure 1. When the knowledge is unknown to the RALM, we set the answer to a refusal expression.
## Experiments
| UE type | UE name | $RGB_{en}$ | $RGB_{zh}$ | | | | | | | | |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| no context | 0p10n | 1p9n | 5p5n | 1p19n | no context | 0p10n | 1p9n | 5p5n | 1p19n | | |
| Verbalize | Verb-1s-1 | 0.445 | 0.139 | 0.208 | 0.023 | 0.042 | 0.477 | 0.441 | 0.119 | 0.242 | 0.124 |
| Verb-1s-5 | 0.253 | 0.186 | 0.182 | 0.160 | 0.179 | 0.173 | 0.170 | 0.182 | 0.170 | 0.198 | |
| Verb-2s-1 | 0.339 | 0.190 | 0.183 | 0.013 | 0.040 | 0.448 | 0.338 | 0.122 | 0.210 | 0.125 | |
| Verb-2s-5 | 0.225 | 0.190 | 0.176 | 0.124 | 0.178 | 0.204 | 0.165 | 0.412 | 0.240 | 0.442 | |
| Consistency | Ent | 0.126 | 0.305 | 0.030 | 0.009 | 0.033 | 0.253 | 0.256 | 0.093 | 0.148 | 0.082 |
| Agree | 0.127 | 0.192 | 0.026 | 0.010 | 0.028 | 0.250 | 0.261 | 0.078 | 0.150 | 0.075 | |
| FSD | 0.104 | 0.162 | 0.041 | 0.014 | 0.048 | 0.201 | 0.182 | 0.083 | 0.122 | 0.086 | |
| Similarity Matrix-based | Eigv | 0.202 | 0.232 | 0.289 | 0.271 | 0.260 | 0.247 | 0.282 | 0.299 | 0.271 | 0.284 |
| Deg | 0.200 | 0.229 | 0.292 | 0.275 | 0.262 | 0.236 | 0.277 | 0.297 | 0.268 | 0.283 | |
Table 1: The Brier score (lower score indicates better calibration) of different UE methods on different RAG settings and datasets.The âApBnâ means A positive chunks and B negative chunks for RAG context settings.
### Experimental Setup
To focus on the modelâs knowledge capacity while minimizing the influence of reasoning, we primarily consider simple factual questions with short answers. These questions typically require only a single evidence document to be answered correctly, for which single-step retrieval is sufficient. Additional details are described in Appendix B.
#### RALM Models
We adopt two prevalent families of open-source LLMs, Qwen and LLaMA. Although modern LLMs are multilingual, We find that Qwen has stronger knowledge in Chinese, whereas LLaMA performs better on English knowledge. To better exploit the knowledge of each model family, we evaluate Qwen https://huggingface.co/Qwen on Chinese datasets and LLaMA https://github.com/meta-llama/llama3 on English datasets. In the main text, we mainly report results for models with approximately 7B parameters. For the retrieval component, document chunks and positive ground-truth passages are provided by the original datasets. We perform hybrid search and re-ranking using Milvus https://milvus.io to construct high-quality negative examples, taking both semantic and lexical similarity into account to provide sufficient difficulty.
#### Hyper-Parameters
The generation temperature is set to 0.5, and the number of sampled generations is set to 16, following (Lyu et al. 2025a). Other generation hyper-parameters are kept at the default values for the corresponding LLMs.
#### Datasets
We explore the RALMsâ performance in open-domain QA tasks, using three prevalent fact-oriented single-hop question datasets to evaluate the performance of LLMs, including two RAG datasets, CRUD (Lyu et al. 2025b) and RGB (Chen et al. 2024), and an QA dataset, NQ (Kwiatkowski et al. 2019). Covering both Chinese and English, the datasets are well-suited for testing Qwen and LLaMA series. NQ and CRUD are large scale QA/RAG datasets suitable for both training and test. RGB is a dataset particular developed for test, including refusal ability of RALMs.
<details>
<summary>x2.png Details</summary>

### Visual Description
## Reliability Diagrams: Qwen-2.5-7B Calibration Across Knowledge Categories and RAG Settings
### Overview
The image displays a 3x4 grid of reliability diagrams (calibration plots) for the Qwen-2.5-7B language model. The diagrams evaluate the model's calibrationâhow well its confidence scores match its actual accuracyâacross four knowledge categories and under three different Retrieval-Augmented Generation (RAG) context settings. Each row corresponds to a specific RAG setting, and each column corresponds to a specific knowledge category.
### Components/Axes
* **Overall Structure:** Three rows labeled (a), (b), and (c). Each row contains four subplots.
* **Subplot Titles (Columns):** From left to right: "highlyknown knowledge", "maybeknown knowledge", "weaklyknown knowledge", "unknown knowledge".
* **Axes (Identical for all subplots):**
* **X-axis:** "Confidence", scaled from 0.0 to 1.0 with increments of 0.2.
* **Y-axis:** "Accuracy", scaled from 0.0 to 1.0 with increments of 0.2.
* **Legend (Top-left of each subplot):**
* **Perfect calibration:** A gray diagonal line from (0,0) to (1,1). Represents ideal calibration where confidence equals accuracy.
* **Accuracy:** Blue bars. The height of each bar represents the model's actual accuracy for predictions within a specific confidence bin.
* **Gap:** Pink bars stacked on top of the blue bars. The total height (blue + pink) represents the model's confidence for that bin. The pink portion visualizes the miscalibration gap (Confidence - Accuracy).
* **Row Captions (Below each row):**
* (a) Reliability Diagram of Qwen-2.5-7B with no context RAG setting
* (b) Reliability Diagram of Qwen-2.5-7B with 0p10n context RAG setting
* (c) Reliability Diagram of Qwen-2.5-7B with 1p9n context RAG setting
### Detailed Analysis
#### Row (a): No Context RAG Setting
* **highlyknown knowledge:** Blue bars (Accuracy) are generally close to or slightly below the perfect calibration line. The pink gaps are small to moderate, indicating reasonable but slightly overconfident calibration.
* **maybeknown knowledge:** Blue bars are consistently and significantly below the perfect calibration line, especially at mid-to-high confidence (0.4-0.8). Large pink gaps indicate substantial overconfidence (confidence > accuracy).
* **weaklyknown knowledge:** Similar to "maybeknown," with blue bars far below the diagonal. Very large pink gaps across most confidence bins show severe overconfidence.
* **unknown knowledge:** Blue bars are extremely low, often near zero, across all confidence levels. The pink gaps are massive, filling almost the entire plot area. This indicates the model is highly overconfident when it lacks knowledge, assigning high confidence to almost entirely incorrect answers.
#### Row (b): 0p10n Context RAG Setting
* **highlyknown knowledge:** Calibration improves compared to (a). Blue bars are closer to the diagonal, with smaller pink gaps, especially at higher confidence levels.
* **maybeknown knowledge:** Improvement is visible. Blue bars are higher than in (a), reducing the pink gap, but a notable overconfidence gap remains, particularly in the 0.6-0.8 confidence range.
* **weaklyknown knowledge:** Shows some improvement over (a). Blue bars are higher, but large pink gaps persist, indicating continued significant overconfidence.
* **unknown knowledge:** Still shows severe miscalibration. While blue bars are slightly higher than in (a) at very high confidence (0.8-1.0), the pink gaps remain enormous, demonstrating the model is still highly overconfident about its unknowns.
#### Row (c): 1p9n Context RAG Setting
* **highlyknown knowledge:** Calibration appears very good. Blue bars closely follow the perfect calibration line with minimal pink gaps.
* **maybeknown knowledge:** Calibration is significantly improved. Blue bars are much closer to the diagonal, with small pink gaps, indicating well-calibrated confidence.
* **weaklyknown knowledge:** Shows the most dramatic improvement. Blue bars are now much higher and closer to the diagonal, though a moderate pink gap remains at the highest confidence bin (0.8-1.0).
* **unknown knowledge:** Improvement is evident but miscalibration persists. Blue bars are higher than in (a) and (b), especially at high confidence, but substantial pink gaps remain, showing the model is still overconfident, though less severely than with no context.
### Key Observations
1. **Knowledge Hierarchy:** Calibration consistently improves from "unknown" to "highlyknown" knowledge across all settings. The model is best calibrated on what it knows well and worst calibrated on what it doesn't know.
2. **RAG Impact:** The context setting has a profound effect on calibration.
* **No Context (a):** Worst calibration, especially severe overconfidence on "unknown" and "weaklyknown" knowledge.
* **0p10n Context (b):** Provides moderate improvement, reducing but not eliminating overconfidence.
* **1p9n Context (c):** Provides the best calibration, bringing "highlyknown" and "maybeknown" close to perfect calibration and significantly improving "weaklyknown."
3. **Persistent Overconfidence on Unknowns:** Even with the best context (1p9n), the model remains notably overconfident in the "unknown knowledge" category, as shown by the persistent pink gaps in the bottom-right subplot.
4. **Visual Trend:** The visual trend across rows (a) -> (b) -> (c) is a reduction in the size of the pink "Gap" bars and a rising of the blue "Accuracy" bars toward the diagonal "Perfect calibration" line, indicating improving calibration with better context.
### Interpretation
These reliability diagrams provide a diagnostic view of the Qwen-2.5-7B model's self-awareness. The data suggests:
1. **Intrinsic Overconfidence:** Without external context (Row a), the model exhibits a strong intrinsic bias toward overconfidence, particularly when operating at the edges of its knowledge ("weaklyknown", "unknown"). This is a common failure mode in LLMs.
2. **Context as a Calibration Tool:** Providing context through RAG acts as a powerful calibration mechanism. The "1p9n" setting (likely meaning 1 positive and 9 negative/irrelevant context passages) appears most effective. It likely helps the model better distinguish between what it knows and what it doesn't, grounding its confidence estimates in external evidence.
3. **Limits of Context:** The persistent gap in the "unknown knowledge" category, even with optimal context, indicates a fundamental challenge. The model may struggle to fully recognize or admit complete ignorance, even when provided with context that doesn't contain the answer. This has implications for reliability in high-stakes applications.
4. **Practical Implication:** For users of this model, the findings argue strongly for using a RAG system, specifically one configured like the "1p9n" setting, to obtain more trustworthy confidence scores. One should be especially skeptical of the model's high-confidence outputs when it is operating on poorly known or unknown topics, as the calibration gap remains significant in those areas.
**Language Declaration:** All text within the image is in English. No other languages are present.
</details>
Figure 3: The reliability diagram under different internal and external knowledge states. The blue bar is the precision questions. The pink bar indicates the over-confident gap, and the purple bar indicates the under-confident gap.
#### Answer Judgment
We first assign a knowledge state to each question based on both temperature-sampled and greedy-decoding results, following (Gekhman et al. 2024). This yields four categories: âhighlyknownâ, âmaybeknownâ, âweaklyknownâ, and âunknownâ. We treat the former two categories as âshould-answerâ and the latter two as âshould-refuseâ according to the precision analysis in Section of RQ1. Following (Sun et al. 2025), we then apply a strict answer-decision workflow to determine whether a model output should be regarded as a refusal or a correct answer, including an LLM-as-a-judge step, exact-match checking, and a refusal-word filter.
#### Evaluation Metrics
Evaluation metrics include accuracy-based and confidence-calibration measures (Feng et al. 2024; Sun et al. 2025). The formal definitions of all metrics are given in Appendix B, and we briefly summarize them as follows:
- Accuracy-based metrics: The answering ability of RALMs is multi-dimensional, reflecting both answer quality and refusal quality.
- Answer Quality (AQ): We report answer precision (Pre), recall (Rec), and F1 for correct answers.
- Refusal Quality (RQ): We measure the refusal rate(RR), refusal precision (RPrec), recall (RRec) and F1(RF1).
- Overall Quality (OQ): We report overall accuracy (OAcc), defined as the proportion of outputs that are either correct answers or proper refusals.
- Confidence calibration metrics: We mainly use Brier Score to measure whether the answer confidence measure the answer precision.
### Do RALMs Know When They Donât Know? (RQ1)
We systematically investigate how prompt variants, positive context position, context quality, and quantity affect the model performance. Detailed discussions are in Appendix C We heuristically varied the numbers of positive and negative examples and examined their impact on the results. In this section, we first examine the calibration error with different UE methods to choose the best one for the following analysis. We then analyze confidence and accuracy in turn as they contribute to the calibration results.
Calibration error of RALMs.
We exclude refusals for UE, since they are outcome-level decisions co-equal with answering, not comparable to specific answer content. Results are in Table 1. The calibration error varies under different RAG settings, and no single method performs best across all scenarios. This aligns with (Moskvoretskii et al. 2025). However, the RALMs become extremely well-calibrated when positive documents exist, especially for verbalize and consistency-based UE methods. This indicates that the UE methods are also acceptable for RALMs. As the consistency-based methods perform best generally, we take their results for further explanation. We contrast the presence versus the absence of context. We find that when no positive context exists (0p10n), the calibration error becomes worse. And when we insert a single positive context (1p9n), the model becomes extremely calibrated. If we insert more positive context (5p5n), the trend of calibration error vary, become better on $RGB_{en}$ and worser on $RGB_{zh}$ . And if we insert more negative context (1p19n), the calibration error does not significantly change. This means that RALMs can sensitively perceive the availability of knowledge. As we find the key factor is the positive context existence, the following settings use 10 context chunks as the default.
<details>
<summary>x3.png Details</summary>

### Visual Description
\n
## Bar Charts: Answer Accuracy and Refusal Rate of Two LLMs on Different Knowledge Levels
### Overview
The image displays a composite figure containing four bar charts arranged in a 2x2 grid. The charts are grouped into two horizontal sections, labeled (a) and (b), each analyzing a different Large Language Model (LLM). Section (a) analyzes "Qwen-2.5-7B" on a dataset or task denoted as "$RGB_{zh}$". Section (b) analyzes "LLaMA-3.1-8B" on "$RGB_{en}$". Each section contains two charts: one for "Answer Accuracy" and one for "Refusal Rate". The charts compare performance across four knowledge levels ("unknown", "weaklyknown", "maybeknown", "highlyknown") under four different context conditions ("no context", "0 pos", "1 pos", "5 pos").
### Components/Axes
**Common Elements Across All Charts:**
* **X-Axis (Categories):** Four discrete conditions: "no context", "0 pos", "1 pos", "5 pos". These likely represent the amount or type of contextual information provided to the model.
* **Legend (Right Side of Each Chart):** Four colored bars representing knowledge levels:
* Blue: `unknown`
* Orange: `weaklyknown`
* Green: `maybeknown`
* Red: `highlyknown`
* **Chart Titles:** Each chart has a title at the top: "Accuracy with 10 context chunks" or "Refusal Rate with 10 chunks".
* **Annotations:** Red arrows are drawn on the charts to highlight specific trends between data points.
**Section (a) - Top Row: Qwen-2.5-7B on $RGB_{zh}$**
* **Left Chart - Answer Accuracy:**
* **Y-Axis:** "Answer Accuracy", scale from 0.0 to 1.0.
* **Right Chart - Refusal Rate:**
* **Y-Axis:** "Refusal Rate", scale from 0.00 to 0.30.
* **Caption (Below charts):** "(a) Answer Accuracy and Refusal Rate of Qwen-2.5-7B on $RGB_{zh}$"
**Section (b) - Bottom Row: LLaMA-3.1-8B on $RGB_{en}$**
* **Left Chart - Answer Accuracy:**
* **Y-Axis:** "Answer Accuracy", scale from 0.0 to 1.0.
* **Right Chart - Refusal Rate:**
* **Y-Axis:** "Refusal Rate", scale from 0.000 to 0.200.
* **Caption (Below charts):** "(b) Answer Accuracy and Refusal Rate of LLaMA-3.1-8B on $RGB_{en}$"
### Detailed Analysis
**Trend Verification & Data Point Extraction (Approximate Values):**
**Section (a) - Qwen-2.5-7B:**
* **Answer Accuracy (Left Chart):**
* **Trend:** Accuracy generally increases for all knowledge levels as the context condition moves from "no context" to "5 pos". The increase is most dramatic for `highlyknown` (red).
* **Data Points (Approx.):**
* `no context`: unknown ~0.0, weaklyknown ~0.02, maybeknown ~0.45, highlyknown ~1.0.
* `0 pos`: unknown ~0.04, weaklyknown ~0.10, maybeknown ~0.12, highlyknown ~0.22.
* `1 pos`: unknown ~0.70, weaklyknown ~0.80, maybeknown ~0.95, highlyknown ~0.98.
* `5 pos`: unknown ~0.70, weaklyknown ~0.82, maybeknown ~0.98, highlyknown ~1.0.
* **Refusal Rate (Right Chart):**
* **Trend:** Refusal rate peaks sharply at the "0 pos" condition for `unknown` and `weaklyknown`, then drops to near zero for "1 pos" and "5 pos". `maybeknown` and `highlyknown` show lower refusal rates overall.
* **Data Points (Approx.):**
* `no context`: unknown ~0.07, weaklyknown ~0.08, maybeknown ~0.02, highlyknown ~0.00.
* `0 pos`: unknown ~0.32, weaklyknown ~0.31, maybeknown ~0.30, highlyknown ~0.25.
* `1 pos`: unknown ~0.01, weaklyknown ~0.01, maybeknown ~0.00, highlyknown ~0.00.
* `5 pos`: All values are ~0.00.
**Section (b) - LLaMA-3.1-8B:**
* **Answer Accuracy (Left Chart):**
* **Trend:** Similar upward trend as context improves. `highlyknown` starts high and remains high. Other categories show significant improvement from "no context"/"0 pos" to "1 pos"/"5 pos".
* **Data Points (Approx.):**
* `no context`: unknown ~0.0, weaklyknown ~0.01, maybeknown ~0.48, highlyknown ~0.99.
* `0 pos`: unknown ~0.20, weaklyknown ~0.14, maybeknown ~0.26, highlyknown ~0.54.
* `1 pos`: unknown ~0.98, weaklyknown ~0.95, maybeknown ~0.94, highlyknown ~0.96.
* `5 pos`: unknown ~0.99, weaklyknown ~0.99, maybeknown ~0.97, highlyknown ~0.99.
* **Refusal Rate (Right Chart):**
* **Trend:** Refusal rate peaks at "0 pos", most prominently for `weaklyknown`. Rates are generally lower than for Qwen-2.5-7B and drop significantly with positive context.
* **Data Points (Approx.):**
* `no context`: unknown ~0.065, weaklyknown ~0.045, maybeknown ~0.010, highlyknown ~0.000.
* `0 pos`: unknown ~0.140, weaklyknown ~0.200, maybeknown ~0.110, highlyknown ~0.040.
* `1 pos`: unknown ~0.030, weaklyknown ~0.005, maybeknown ~0.012, highlyknown ~0.020.
* `5 pos`: unknown ~0.020, weaklyknown ~0.005, maybeknown ~0.005, highlyknown ~0.015.
### Key Observations
1. **Context is Critical:** For both models, providing positive context ("1 pos", "5 pos") dramatically increases answer accuracy and reduces refusal rates compared to "no context" or "0 pos" conditions.
2. **The "0 pos" Problem:** The "0 pos" condition (likely meaning zero positive examples in context) causes a severe performance drop, especially for less-known items (`unknown`, `weaklyknown`), where accuracy plummets and refusal rates spike.
3. **Knowledge Level Hierarchy:** There is a clear performance hierarchy: `highlyknown` > `maybeknown` > `weaklyknown` > `unknown`. This holds across most conditions, with the gap narrowing significantly when ample positive context ("1 pos", "5 pos") is provided.
4. **Model Comparison:** LLaMA-3.1-8B (on $RGB_{en}$) appears to achieve higher accuracy with "1 pos" context for the `unknown` category compared to Qwen-2.5-7B (on $RGB_{zh}$). Its refusal rates are also generally lower.
### Interpretation
The data demonstrates a strong interaction between a model's pre-existing knowledge about a topic and the quality of the contextual information provided. The key finding is that **context is not universally beneficial**. The "0 pos" condition acts as a detrimental interference, likely confusing the model or triggering conservative refusal behaviors, particularly for information it is uncertain about (`unknown`, `weaklyknown`).
The charts suggest that for these models to perform well, they need either:
a) Strong pre-existing knowledge (`highlyknown`), or
b) High-quality, positive contextual examples ("1 pos" or "5 pos").
The absence of positive examples ("0 pos") is worse than having no context at all for lower-knowledge items. This has practical implications for prompt engineering and retrieval-augmented generation (RAG) systems: ensuring retrieved context contains relevant, positive examples is crucial, as presenting irrelevant or zero-shot examples can actively harm performance. The difference between the two models' results on $RGB_{zh}$ and $RGB_{en}$ may also point to variations in training data, model architecture, or the inherent difficulty of the tasks in different languages.
</details>
Figure 4: The answer precision (denoted as âaccuracyâ) and refusal rate vary according to the internal/external knowledge states. The whole negative context (0 pos) leads to significant decrease of accuracy and increase of refusal on âhighlyknownâ questions.
Over-confident or under-confident.
In this section, we examine how confidence scores vary, given that base LLMs are known to be over-confident (Li et al. 2025) as shown in Figure 3. In the no-context setting, the âhighlyknownâ type is slightly under-confident, whereas the other types are over-confident. The âhighlyknownâ questions attain relatively high confidence values, while the confidence of the other types is more dispersed. However, in the all-negative-context setting, the RALMs become strongly over-confident and the confidence scores for all types become highly dispersed. For âhighlyknownâ questions, the LLM could answer correctly without retrieval, yet the observed accuracy is noticeably worse. This indicates that both accuracy and confidence are substantially affected by noisy contexts. Interestingly, âweaklyknownâ questions achieve higher accuracy under negative contexts, suggesting that the injected noise can have unexpected effects. This finding is consistent with Cuconasu et al. (2024), while we further delineate how this effect depends on specific knowledge categories. Finally, even when one positive context is provided, RALMs tend to be under-confident for most knowledge types, except for the âunknownâ category. Across knowledge types, the model attains high accuracy and more concentrated confidence distributions, indicating that RALMs can effectively detect and exploit helpful information. In summary, these observations explain the calibration trends in Table 1: with all-negative context, accuracy generally decreases and confidence becomes more diffuse, whereas with positive context, accuracy improves and confidence becomes more concentrated.
| RALMs test setting | Method name | CalErr | OQ | AQ | RQ | | | | | | | |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| OaBs $(\downarrow)$ | OAcc $(\uparrow)$ | Pre $(\uparrow)$ | Rec $(\uparrow)$ | F1 $(\uparrow)$ | MA $(\downarrow)$ | RR | OR $(\downarrow)$ | RPre $(\uparrow)$ | RRec $(\uparrow)$ | RF1 $(\uparrow)$ | | |
| Qwen-2.5-7B | | | | | | | | | | | | |
| no context | Vanilla | 0.245 | 0.427 | 0.411 | 1.000 | 0.583 | 0.217 | 0.027 | 0.000 | 1.000 | 0.044 | 0.085 |
| R-tuning | 0.191 | 0.457 | 0.395 | 0.857 | 0.541 | 0.336 | 0.190 | 0.105 | 0.719 | 0.218 | 0.335 | |
| ICFT (n) | 0.226 | 0.487 | 0.450 | 0.953 | 0.611 | 0.250 | 0.103 | 0.039 | 0.806 | 0.145 | 0.245 | |
| ICFT (p) | 0.169 | 0.443 | 0.443 | 1.000 | 0.614 | 0.250 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | |
| ICFT (pn) | 0.167 | 0.440 | 0.440 | 1.000 | 0.611 | 0.243 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | |
| ICFT (w) | 0.181 | 0.423 | 0.414 | 1.000 | 0.585 | 0.296 | 0.017 | 0.000 | 1.000 | 0.028 | 0.055 | |
| 0p10n | Vanilla | 0.325 | 0.290 | 0.168 | 0.372 | 0.231 | 0.500 | 0.363 | 0.355 | 0.505 | 0.257 | 0.341 |
| R-tuning | 0.408 | 0.457 | 0.294 | 0.195 | 0.235 | 0.184 | 0.717 | 0.678 | 0.521 | 0.651 | 0.579 | |
| ICFT (n) | 0.216 | 0.620 | 0.578 | 0.709 | 0.637 | 0.158 | 0.423 | 0.270 | 0.677 | 0.541 | 0.601 | |
| ICFT (p) | 0.204 | 0.400 | 0.400 | 1.000 | 0.571 | 0.342 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | |
| ICFT (pn) | 0.189 | 0.430 | 0.430 | 1.000 | 0.601 | 0.309 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | |
| ICFT (w) | 0.217 | 0.460 | 0.436 | 0.976 | 0.603 | 0.296 | 0.060 | 0.020 | 0.833 | 0.086 | 0.156 | |
| 1p9n | Vanilla | 0.079 | 0.863 | 0.863 | 1.000 | 0.927 | 0.013 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 |
| R-tuning | 0.127 | 0.830 | 0.853 | 0.960 | 0.903 | 0.033 | 0.070 | 0.066 | 0.524 | 0.212 | 0.301 | |
| ICFT (n) | 0.164 | 0.787 | 0.835 | 0.881 | 0.858 | 0.033 | 0.230 | 0.171 | 0.623 | 0.531 | 0.573 | |
| ICFT (p) | 0.068 | 0.827 | 0.827 | 1.000 | 0.905 | 0.072 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | |
| ICFT (pn) | 0.085 | 0.820 | 0.820 | 1.000 | 0.901 | 0.059 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | |
| ICFT (w) | 0.094 | 0.827 | 0.827 | 1.000 | 0.905 | 0.053 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | |
Table 2: Evaluation of refusal trained models under different settings. $(\uparrow)$ indicates a higher score is better, and $(\downarrow)$ vice versa. If no arrow is marked, then the score have no directionality. The best result under a RALMs test settings is marked bold and we do not mark those â1.000â scores. The over-refusal score (OR) which is marked in red indicates the worst case.
Precision and refusal rate.
We begin by analyzing how answer correctness varies. In the all-negative(0 pos) setting, we observe a decline on âhighlyknownâ and âmaybeknownâ questions and a gain on âweaklyknownâ and âunknownâ ones compared to the no-context setting. When a positive context exists, the precision significantly increases, especially for unknown and weakly known knowledge. Increasing the count of positives yields no significant gains in precision. This indicates that LLMs are sensitive to both harmful and supportive contexts. While increasing the number of positive and negative examples does not significantly alter the modelâs response for fact-oriented questions in this kind of shorter context. Then we analyze refusal rate. In the all-negative (0 pos) setting, we observe an significant increase on all the knowledge types. Considering the LLMs can correctly answer âhighlyknownâ questions on their own, refusal on those questions are not correct. We identify this phenomenon as over-refusal, which are not observed in previously research. Likewise, the presence of positive chunk markedly reduces refusal. This is consistent with the pattern of accuracy changes.
Summary.
In this section, we empirically show that RALMs generally âknow they donât knowâ under no-context and positive-context settings. However, they become over-confident when confronted with negative context and may over-refuse questions whose answers they actually know.
### How does RALMsâ refusal ability align with its calibration quality? (RQ2)
| Method name | DR | CU | | |
| --- | --- | --- | --- | --- |
| no context | 0p10n | 10p0n | 1p9n | |
| Vanilla | 0.579 | 0.191 | 0.759 | 0.738 |
| R-tuning | 0.444 | 0.138 | 0.750 | 0.682 |
| ICFT (n) | 0.734 | 0.632 | 0.750 | 0.591 |
| ICFT (p) | 0.750 | 0.658 | 0.824 | 0.723 |
| ICFT (pn) | 0.757 | 0.691 | 0.777 | 0.696 |
| ICFT (w) | 0.704 | 0.684 | 0.770 | 0.703 |
Table 3: Results of denoise rate and positive context utilization.
We adjust refusal ability though the R-tuning and In-context Fine-tuning variants. Considering the knowledge quadrants of Figure 1, we set four ICFT variants as follows:
- ICFT(n) : We append only negative contexts for LLMs, thus the answer of training samples depend on the internal state of LLMs. If internal knowledge entail the question, the answer is original ground truth; else the answer is âI donât knownâ.
- ICFT(p) : We append only positive contexts for LLMs. The answers are all set to original ground truth.
- ICFT(pn): We append both positive and negative contexts for LLMs and the answers are all set to original ground truth. This is because the LLMs can distinguish the positive context and we want to enhance this ability.
- ICFT(w): We include both the ICFT(n) and ICFT(pn) training samples.
We use the training query, only different context and answers to ensure the training fairness. Training and model selection details are in Appendix D We also test RL-based refusal-aware methods.
Response quality of RIFT models
The response quality of refusal-trained RALMs is multi-dimensional. As shown in Table 2, model performance varies across different RALM settings. In the no-context setting, ICFT(n) achieves the best overall accuracy (OAcc, OQ), while ICFT(p) performs best in terms of F1 (AQ). The R-tuning model obtains the highest RF1 (RQ), with ICFT(n) ranking second. This may be because the R-tuning training scenario closely matches the test setting, leading to a higher refusal rate (RR) and moderate refusal precision (RPrec). However, the over-refusal rate (OR) also increases, suggesting that R-tuning may harm the modelâs self-awareness. The decrease in answer precision (Pre) and the increase in mis-answer rate (MR) support this finding. We will further examine the corresponding change in confidence calibration in the following subsection. In the all-negative (0p10n) setting, ICFT(n) performs substantially better than the other models in terms of OAcc (OQ), F1 (OQ), and RF1 (RQ). Although the over-refusal rate (OR) of R-tuning is the worst, ICFT(n) alleviates this issue and performs better than the vanilla RALMs. Moreover, we find that ICFT variants with positive context substantially reduce over-refusal while maintaining competitive overall accuracy (OAcc, OQ). Surprisingly, when positive context is available, the vanilla RALMs achieve the best OAcc (OQ) and F1 (AQ). From the perspective of RQ, ICFT(n) actually appears to perform the best. However, we emphasize that RQ in this positive-context setting should be interpreted with caution, as we do not relabel the âshould-answerâ set in order to remain consistent with the previous two settings.
Refusal Confidence of RIFT models
In RQ1 we do not consider the refusal part, we check the overall brier score (OaBs) as in Table 2. We notice that the performance of calibration error do not align with overall,answer, or refusal quality. Surprisingly, ICFT with positive context(p/pn) get best calibration performance, though their refusal performance is not good as ICFT(n). This provides support for jointly considering active and passive refusals. We provide confidence distribution illustration Appendix D.
Retrieval handling of RIFT models
Because a single calibration-error metric cannot fully reflect refusal quality, we introduce retrieval-handling metrics to further explain the results. Intuitively, a model that is more robust to noise is more likely to rely on its internal knowledge. While some methods (Zhang et al. 2025; Bi et al. 2025) explicitly emphasize the context faithfulness of RALMs. We evaluate these abilities using the denoising rate (DR) and the context utilization rate (CU), as reported in Table 3. In terms of denoising ability, all ICFT models perform better than the vanilla models, whereas the R-tuning models perform worse than the vanilla baseline. Although the R-tuning methods outperform the vanilla models in OAcc (OQ) and RF1 (RQ), this suggests that R-tuning primarily encourages models to refuse based on their internal states rather than to resist noisy context. However, the R-tuning approach appears to sacrifice the modelâs underlying knowledge competence in exchange for a stronger ability to articulate refusals, according to its worse DR performance in no context settings. In terms of context utilization, we find that ICFT(p) yields better results, while including negative context leads to worse performance in the all-positive (10p) setting. Surprisingly, however, all refusal fine-tuned models perform worse than the vanilla RALMs. This explains why these models perform poorly in scenarios with positive evidence: they tend to refuse internally unknown questions while ignoring the positive context.
Summary
In this section, our results show that the over-refusal problem is mitigated by In-context fine-tuning, but magnified by R-tuning. The systemâs performance should be assessed by jointly considering the modelâs confidence, robustness, and context faithfulness. However, we also find that the refusal ability may conflict with the quality of the answer.
### Mitigating the Over-refusal Issue in RALMs (RQ3)
Although some refusal-aware RALM models do not support appropriate abstention by themselves, their confidence profiles can still distinguish correct refusals from incorrect ones. To validate whether we can distinguish different knowledge states and enable more appropriate refusals, we first study a simple threshold-based post-refusal technique. Concretely, we follow the thresholds-based refusal at inference stage.
To reduce the negative effects introduced by noisy contexts, we further develop a two-stage refusal technique. In the first stage, we apply a threshold $T_{s}$ to $U_{\text{LLM}}$ (the uncertainty of the base LLM) to detect whether the answer can be supported by internal knowledge, and a threshold on $\Delta U=U_{\text{RALM}}-U_{\text{LLM}}$ (where $U_{\text{RALM}}$ is the uncertainty of the RALM, which incorporates context) to infer the knowledge state. In the second stage, we apply a refusal threshold in the same way as the baseline, but only when the RALM is classified as âunknownâ. All threshold values are selected via grid search on the development set. To better isolate the effect of knowledge on refusal, we compare these methods under an idealized but challenging (0p10n) context configuration. The results are summarized in Table 4. The post-refusal methods achieve higher overall accuracy than their counterparts in Table 2, but they also exhibit a substantially higher over-refusal rate. By first determining the knowledge state of the LLM itself, the model can choose when to rely on its own knowledge, yielding more calibrated confidence estimates and enabling further refusals without overusing harmful negative contexts, especially for ICFT(p) which show better calibration but less tendency to refuse on its own. Finally, we note that Wang et al. (2025) adopts similar information-gain-based method to detect context utility. This further supports our findings, while we provide a more explicit analysis of how knowledge states influence refusal behavior. Additional details for real RAG experiments are provided in Appendix E.
| Refusal method | Method Name | OQ | AQ | RQ | | |
| --- | --- | --- | --- | --- | --- | --- |
| OAcc | MA | AF1 | OR | RF1 | | |
| 0p10n | | | | | | |
| Post refusal | Vanilla | 0.437 | 0.145 | 0.167 | 0.770 | 0.570 |
| ICFT(n) | 0.673 | 0.098 | 0.655 | 0.462 | 0.690 | |
| ICFT(p) | 0.683 | 0.240 | 0.672 | 0.243 | 0.682 | |
| Ours | Vanilla | 0.523 | 0.104 | 0.240 | 0.282 | 0.590 |
| ICFT(n) | 0.729 | 0.059 | 0.707 | 0.176 | 0.731 | |
| ICFT(p) | 0.697 | 0.178 | 0.691 | 0.106 | 0.698 | |
Table 4: RALMs knowledge state aware refusal technique.
## Conclusions
In this work, we investigate whether RALMs âknow when they donât knowâ. We find that the calibration state of RALMs is greatly influenced by external contexts. In particular, we identify that purely negative contexts severely harm calibration and induce an over-refusal problem. We further study how the refusal quality of RALMs aligns with their calibration and observe that refusal-aware RALMs struggle to handle different RAG settings, due to entangled internal knowledge states and reduced context utilization. Finally, we combine the refusal ability of LLMs with post-refusal methods to balance overall response quality while mitigating over-refusal. Our study offers insights that underscore the need for improved calibration methods and the explicit modeling of dynamically evolving knowledge.
## Acknowledgments
The authors thank all the reviewers for their suggestions and comments. This work is supported by National Natural Science Foundation of China (No.U21B2009). It is also supported by scholarship under the State Scholarship Fund and a visiting to Singapore Management University organized by the China Scholarship Council (CSC). The authors also acknowledge the material support by Boston Meditech Group and Hangzhou Kangyi Health Management Limited Partnership.
## References
- Asai et al. (2024) Asai, A.; Wu, Z.; Wang, Y.; Sil, A.; and Hajishirzi, H. 2024. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. In The Twelfth International Conference on Learning Representations.
- Bi et al. (2025) Bi, B.; Huang, S.; Wang, Y.; Yang, T.; Zhang, Z.; Huang, H.; Mei, L.; Fang, J.; Li, Z.; Wei, F.; et al. 2025. Context-dpo: Aligning language models for context-faithfulness. In Findings of the Association for Computational Linguistics: ACL 2025, 10280â10300.
- Chen et al. (2024) Chen, J.; Lin, H.; Han, X.; and Sun, L. 2024. Benchmarking large language models in retrieval-augmented generation. In Proceedings of the AAAI Conference on Artificial Intelligence, 17754â17762.
- Cheng et al. (2024) Cheng, Q.; Sun, T.; Liu, X.; Zhang, W.; Yin, Z.; Li, S.; Li, L.; He, Z.; Chen, K.; and Qiu, X. 2024. Can AI Assistants Know What They Donât Know? In International Conference on Machine Learning, 8184â8202. PMLR.
- Cuconasu et al. (2024) Cuconasu, F.; Trappolini, G.; Siciliano, F.; Filice, S.; Campagnano, C.; Maarek, Y.; Tonellotto, N.; and Silvestri, F. 2024. The power of noise: Redefining retrieval for rag systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, 719â729.
- Deng et al. (2025) Deng, Y.; Li, M.; Pang, L.; Zhang, W.; and Lam, W. 2025. Unveiling Knowledge Boundary of Large Language Models for Trustworthy Information Access. In SIGIR 2025, 4086â4089. ACM.
- Deng et al. (2024) Deng, Y.; Zhao, Y.; Li, M.; Ng, S. K.; and Chua, T.-S. 2024. Donât Just Say âI donât knowâ! Self-aligning Large Language Models for Responding to Unknown Questions with Explanations. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 13652â13673.
- Fang et al. (2024) Fang, F.; Bai, Y.; Ni, S.; Yang, M.; Chen, X.; and Xu, R. 2024. Enhancing Noise Robustness of Retrieval-Augmented Language Models with Adaptive Adversarial Training. In ACL (1).
- Feng et al. (2024) Feng, S.; Shi, W.; Wang, Y.; Ding, W.; Balachandran, V.; and Tsvetkov, Y. 2024. Donât Hallucinate, Abstain: Identifying LLM Knowledge Gaps via Multi-LLM Collaboration. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 14664â14690.
- Gekhman et al. (2024) Gekhman, Z.; Yona, G.; Aharoni, R.; Eyal, M.; Feder, A.; Reichart, R.; and Herzig, J. 2024. Does Fine-Tuning LLMs on New Knowledge Encourage Hallucinations? In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 7765â7784.
- Geng et al. (2024) Geng, J.; Cai, F.; Wang, Y.; Koeppl, H.; Nakov, P.; and Gurevych, I. 2024. A Survey of Confidence Estimation and Calibration in Large Language Models. In Duh, K.; GĂłmez-Adorno, H.; and Bethard, S., eds., Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), NAACL 2024, Mexico City, Mexico, June 16-21, 2024, 6577â6595. Association for Computational Linguistics.
- Huang et al. (2025) Huang, L.; Yu, W.; Ma, W.; Zhong, W.; Feng, Z.; Wang, H.; Chen, Q.; Peng, W.; Feng, X.; Qin, B.; and Liu, T. 2025. A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions. ACM Trans. Inf. Syst., 43(2): 42:1â42:55.
- Jeong et al. (2024) Jeong, S.; Baek, J.; Cho, S.; Hwang, S. J.; and Park, J. C. 2024. Adaptive-RAG: Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 7029â7043.
- Kadavath et al. (2022) Kadavath, S.; Conerly, T.; Askell, A.; Henighan, T.; Drain, D.; Perez, E.; Schiefer, N.; Hatfield-Dodds, Z.; DasSarma, N.; Tran-Johnson, E.; Johnston, S.; Showk, S. E.; Jones, A.; Elhage, N.; Hume, T.; Chen, A.; Bai, Y.; Bowman, S.; Fort, S.; Ganguli, D.; Hernandez, D.; Jacobson, J.; Kernion, J.; Kravec, S.; Lovitt, L.; Ndousse, K.; Olsson, C.; Ringer, S.; Amodei, D.; Brown, T.; Clark, J.; Joseph, N.; Mann, B.; McCandlish, S.; Olah, C.; and Kaplan, J. 2022. Language Models (Mostly) Know What They Know. CoRR, abs/2207.05221.
- Kapoor et al. (2024) Kapoor, S.; Gruver, N.; Roberts, M.; Collins, K.; Pal, A.; Bhatt, U.; Weller, A.; Dooley, S.; Goldblum, M.; and Wilson, A. G. 2024. Large language models must be taught to know what they donât know. Advances in Neural Information Processing Systems, 37: 85932â85972.
- Kwiatkowski et al. (2019) Kwiatkowski, T.; Palomaki, J.; Redfield, O.; Collins, M.; Parikh, A.; Alberti, C.; Epstein, D.; Polosukhin, I.; Devlin, J.; Lee, K.; et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7: 453â466.
- Lee, Lin, and Tan (2025) Lee, Z. P.; Lin, A.; and Tan, C. 2025. Finetune-RAG: Fine-Tuning Language Models to Resist Hallucination in Retrieval-Augmented Generation. arXiv preprint arXiv:2505.10792.
- Lewis et al. (2020) Lewis, P.; Perez, E.; Piktus, A.; Petroni, F.; Karpukhin, V.; Goyal, N.; KĂŒttler, H.; Lewis, M.; Yih, W.; RocktĂ€schel, T.; Riedel, S.; and Kiela, D. 2020. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In Larochelle, H.; Ranzato, M.; Hadsell, R.; Balcan, M.; and Lin, H., eds., Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.
- Li et al. (2025) Li, M.; Zhao, Y.; Zhang, W.; Li, S.; Xie, W.; Ng, S.; Chua, T.; and Deng, Y. 2025. Knowledge Boundary of Large Language Models: A Survey. In ACL 2025.
- Li et al. (2024) Li, Z.; Hu, X.; Liu, A.; Zheng, K.; Huang, S.; and Xiong, H. 2024. Refiner: Restructure Retrieved Content Efficiently to Advance Question-Answering Capabilities. In Findings of the Association for Computational Linguistics: EMNLP 2024, 8548â8572.
- Lin, Trivedi, and Sun (2024) Lin, Z.; Trivedi, S.; and Sun, J. 2024. Generating with Confidence: Uncertainty Quantification for Black-box Large Language Models. Trans. Mach. Learn. Res., 2024.
- Lyu et al. (2025a) Lyu, Q.; Shridhar, K.; Malaviya, C.; Zhang, L.; Elazar, Y.; Tandon, N.; Apidianaki, M.; Sachan, M.; and Callison-Burch, C. 2025a. Calibrating large language models with sample consistency. In Proceedings of the AAAI Conference on Artificial Intelligence, 19260â19268.
- Lyu et al. (2025b) Lyu, Y.; Li, Z.; Niu, S.; Xiong, F.; Tang, B.; Wang, W.; Wu, H.; Liu, H.; Xu, T.; and Chen, E. 2025b. Crud-rag: A comprehensive chinese benchmark for retrieval-augmented generation of large language models. ACM Transactions on Information Systems, 43(2): 1â32.
- Mei et al. (2025) Mei, Z.; Zhang, C.; Yin, T.; Lidard, J.; Shorinwa, O.; and Majumdar, A. 2025. Reasoning about Uncertainty: Do Reasoning Models Know When They Donât Know? arXiv preprint arXiv:2506.18183.
- Moskvoretskii et al. (2025) Moskvoretskii, V.; Marina, M.; Salnikov, M.; Ivanov, N.; Pletenev, S.; Galimzianova, D.; Krayko, N.; Konovalov, V.; Nikishina, I.; and Panchenko, A. 2025. Adaptive retrieval without self-knowledge? bringing uncertainty back home. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 6355â6384.
- Park and Lee (2024) Park, S.-I.; and Lee, J.-Y. 2024. Toward robust ralms: Revealing the impact of imperfect retrieval on retrieval-augmented language models. Transactions of the Association for Computational Linguistics, 12: 1686â1702.
- Ram et al. (2023) Ram, O.; Levine, Y.; Dalmedigos, I.; Muhlgay, D.; Shashua, A.; Leyton-Brown, K.; and Shoham, Y. 2023. In-context retrieval-augmented language models. Transactions of the Association for Computational Linguistics, 11: 1316â1331.
- Soudani, Zamani, and Hasibi (2025) Soudani, H.; Zamani, H.; and Hasibi, F. 2025. Uncertainty Quantification for Retrieval-Augmented Reasoning. arXiv preprint arXiv:2510.11483.
- Su et al. (2024) Su, W.; Tang, Y.; Ai, Q.; Wu, Z.; and Liu, Y. 2024. DRAGIN: Dynamic Retrieval Augmented Generation based on the Real-time Information Needs of Large Language Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 12991â13013.
- Sun et al. (2025) Sun, X.; Xie, J.; Chen, Z.; Liu, Q.; Wu, S.; Chen, Y.; Song, B.; Wang, Z.; Wang, W.; and Wang, L. 2025. Divide-then-align: Honest alignment based on the knowledge boundary of rag. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 11461â11480.
- Tian et al. (2023) Tian, K.; Mitchell, E.; Zhou, A.; Sharma, A.; Rafailov, R.; Yao, H.; Finn, C.; and Manning, C. D. 2023. Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 5433â5442.
- Wang et al. (2025) Wang, Z.; Liang, Z.; Shao, Z.; Ma, Y.; Dai, H.; Chen, B.; Mao, L.; Lei, C.; Ding, Y.; and Li, H. 2025. InfoGain-RAG: Boosting Retrieval-Augmented Generation through Document Information Gain-based Reranking and Filtering. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 7201â7215.
- Wei, Chen, and Meng (2024) Wei, Z.; Chen, W.-L.; and Meng, Y. 2024. Instructrag: Instructing retrieval-augmented generation via self-synthesized rationales. arXiv preprint arXiv:2406.13629.
- Wen et al. (2025) Wen, B.; Yao, J.; Feng, S.; Xu, C.; Tsvetkov, Y.; Howe, B.; and Wang, L. L. 2025. Know your limits: A survey of abstention in large language models. Transactions of the Association for Computational Linguistics, 13: 529â556.
- Xu, Shi, and Choi (2024) Xu, F.; Shi, W.; and Choi, E. 2024. RECOMP: Improving Retrieval-Augmented LMs with Context Compression and Selective Augmentation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net.
- Yin et al. (2023) Yin, Z.; Sun, Q.; Guo, Q.; Wu, J.; Qiu, X.; and Huang, X.-J. 2023. Do Large Language Models Know What They Donât Know? In Findings of the Association for Computational Linguistics: ACL 2023, 8653â8665.
- Yoran et al. (2024) Yoran, O.; Wolfson, T.; Ram, O.; and Berant, J. 2024. Making Retrieval-Augmented Language Models Robust to Irrelevant Context. In The Twelfth International Conference on Learning Representations.
- Zhang et al. (2024) Zhang, H.; Diao, S.; Lin, Y.; Fung, Y.; Lian, Q.; Wang, X.; Chen, Y.; Ji, H.; and Zhang, T. 2024. R-tuning: Instructing large language models to say âi donât knowâ. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 7106â7132.
- Zhang et al. (2025) Zhang, Q.; Xiang, Z.; Xiao, Y.; Wang, L.; Li, J.; Wang, X.; and Su, J. 2025. FaithfulRAG: Fact-Level Conflict Modeling for Context-Faithful Retrieval-Augmented Generation. In Che, W.; Nabende, J.; Shutova, E.; and Pilehvar, M. T., eds., Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 21863â21882. Vienna, Austria: Association for Computational Linguistics. ISBN 979-8-89176-251-0.
- Zhu et al. (2025) Zhu, R.; Ma, Z.; Wu, J.; Gao, J.; Wang, J.; Lin, D.; and He, C. 2025. Utilize the flow before stepping into the same river twice: Certainty represented knowledge flow for refusal-aware instruction tuning. In Proceedings of the AAAI Conference on Artificial Intelligence, 26157â26165.
- Zhu, Panigrahi, and Arora (2025) Zhu, X.; Panigrahi, A.; and Arora, S. 2025. On the power of context-enhanced learning in llms. arXiv preprint arXiv:2503.01821.