## Chart/Diagram Type: Performance Comparison Chart ### Overview The image presents a performance comparison of different language models on various tasks. It includes a radar chart comparing "AgentFlow" and "AgentFlow (w/o Flow-GRPO)" across several tasks, and bar charts showing the performance of different models on specific tasks like 2Wiki (Search), HotpotQA (Search), GAIA (Agentic), AIME24 (Math), GameOf24 (Math), and GPQA (Science). The performance metric used is accuracy (%). ### Components/Axes **Radar Chart:** * **Title:** Comparison of AgentFlow and AgentFlow (w/o Flow-GRPO) * **Data Series:** * AgentFlow (Red Line) * AgentFlow (w/o Flow-GRPO) (Blue Line) * **Categories (Spokes):** * Bamboogle * MedQA * GPQA * GameOf24 * AMC23 * AIME24 * GAIA * Musique * HotpotQA * 2Wiki * **Values:** Accuracy scores are plotted along each spoke. **Bar Charts:** * **Y-axis:** Accuracy (%) ranging from 0 to 80 (varies by chart). * **X-axis:** Different language models. * **Legend (located below the bar charts):** * Qwen-2.5-7B (Light Gray) * TIR (7B) (Light Green) * Search-R1 (7B) (Light Blue) * AutoGen (7B) (Light Purple) * GPT-4o (~200B) (Green) * ToRL (7B) (Blue) * ReSearch (7B) (Purple) * AgentFlow (7B) (Red) ### Detailed Analysis **Radar Chart:** * **AgentFlow (Red):** * Bamboogle: 69.6 * MedQA: 80.0 * GPQA: 47.0 * GameOf24: 53.0 * AMC23: 61.5 * AIME24: 40.0 * GAIA: 33.1 * Musique: 25.3 * HotpotQA: 57.0 * 2Wiki: 71.2 * **AgentFlow (w/o Flow-GRPO) (Blue):** * Bamboogle: 58.4 * MedQA: 76.0 * GPQA: 37.0 * GameOf24: 31.0 * AMC23: 47.4 * AIME24: 16.7 * GAIA: 17.2 * Musique: 19.2 * HotpotQA: 51.3 * 2Wiki: 60.0 **Bar Charts:** * **2Wiki (Search):** * Qwen-2.5-7B: 23.0% * GPT-4o (~200B): 49.5% * Search-R1 (7B): 38.2% * ToRL (7B): 47.6% * AutoGen (7B): 44.0% * AgentFlow (7B): 77.2% * **HotpotQA (Search):** * Qwen-2.5-7B: 21.0% * GPT-4o (~200B): 54.0% * Search-R1 (7B): 37.0% * ToRL (7B): 43.5% * AutoGen (7B): 50.0% * AgentFlow (7B): 57.0% * **GAIA (Agentic):** * Qwen-2.5-7B: 3.2% * GPT-4o (~200B): 17.3% * Search-R1 (7B): 19.1% * ToRL (7B): 17.3% * AutoGen (7B): 6.3% * AgentFlow (7B): 33.1% * **AIME24 (Math):** * Qwen-2.5-7B: 6.7% * GPT-4o (~200B): 13.3% * Search-R1 (7B): 10.0% * ToRL (7B): 13.3% * AgentFlow (7B): 40.0% * **GameOf24 (Math):** * Qwen-2.5-7B: 33.0% * GPT-4o (~200B): 32.0% * TIR (7B): 33.0% * ToRL (7B): 31.0% * AgentFlow (7B): 53.0% * AutoGen (7B): 24.0% * **GPQA (Science):** * Qwen-2.5-7B: 34.0% * GPT-4o (~200B): 42.0% * TIR (7B): 31.0% * ToRL (7B): 42.0% * AutoGen (7B): 35.0% * AgentFlow (7B): 47.0% ### Key Observations * AgentFlow consistently outperforms AgentFlow (w/o Flow-GRPO) across all tasks in the radar chart. * AgentFlow (7B) significantly outperforms other models in 2Wiki (Search), HotpotQA (Search), GAIA (Agentic), AIME24 (Math), and GameOf24 (Math). * GPT-4o (~200B) shows competitive performance, often being the second-best performing model. * The performance of different models varies significantly across different tasks. ### Interpretation The data suggests that AgentFlow benefits significantly from the "Flow-GRPO" component, as evidenced by its superior performance compared to the version without it. AgentFlow (7B) demonstrates strong capabilities across a diverse set of tasks, indicating its potential as a versatile language model. The performance differences between models highlight the importance of model architecture and training data for specific tasks. The radar chart provides a holistic view of AgentFlow's strengths and weaknesses relative to its variant, while the bar charts offer a detailed comparison against other models on individual tasks. The "+X.X%" annotations near the radar chart indicate the percentage improvement of AgentFlow over AgentFlow (w/o Flow-GRPO) for specific categories (Science, Search, Math, Agentic).

## Mixed Chart Analysis: AgentFlow Performance Across Diverse Tasks ### Overview The image presents a comparative analysis of "AgentFlow" performance against a baseline "AgentFlow (w/o Flow-GRPO)" and several other models across various tasks. The left side features a radar chart illustrating the relative performance of two AgentFlow variants across ten distinct tasks, grouped into four broader categories (Science, Search, Agentic, Math), with aggregated percentage improvements. The right side displays six individual bar charts, providing a more granular comparison of "AgentFlow (7B)" against up to seven other models (Qwen-2.5-7B, GPT-4o (~200B), TIR (7B), ToRL (7B), Search-R1 (7B), ReSearch (7B), AutoGen (7B)) for specific tasks, showing "Accuracy (%)". ### Components/Axes **Left Side: Radar Chart** * **Title:** Implicitly comparing "AgentFlow (w/o Flow-GRPO)" and "AgentFlow". * **Legend (Top-Left):** * Blue line with circular markers: "AgentFlow (w/o Flow-GRPO)" * Red line with circular markers: "AgentFlow" * **Radial Axes/Categories (Clockwise from top):** * Bamboogle * 2Wiki * HotpotQA * Musique * GAIA * AIME24 * AMC23 * GameOf24 * GPQA * MedQA * **Radial Scale:** Concentric circles represent increasing performance values, likely percentages. The innermost circle represents 0, and the outermost visible circle corresponds to values up to 80.0. * **Additional Labels:** * Top-left, near MedQA: "+7.0% Science" * Top-right, near 2Wiki: "+10.1% Search" * Bottom-right, near GAIA: "+15.9% Agentic" * Bottom-left, near GameOf24: "+19.8% Math" **Right Side: Bar Charts** * **Common Y-axis (Left side of each bar chart):** "Accuracy (%)" * **Common Legend (Positioned centrally, below the top row of bar charts):** * Light Gray bar: Qwen-2.5-7B * Dark Gray bar: GPT-4o (~200B) * Light Green bar: TIR (7B) * Light Blue bar: ToRL (7B) * Medium Blue bar: Search-R1 (7B) * Dark Blue bar: ReSearch (7B) * Light Purple bar: AutoGen (7B) * Red bar: AgentFlow (7B) * **Individual Chart Titles:** * Top-left: 2Wiki (Search) * Top-middle: HotpotQA (Search) * Top-right: GAIA (Agentic) * Bottom-left: AIME24 (Math) * Bottom-middle: GameOf24 (Math) * Bottom-right: GPQA (Science) ### Detailed Analysis **Radar Chart (Left Side)** The radar chart visually compares the performance of "AgentFlow" (red line) against "AgentFlow (w/o Flow-GRPO)" (blue line) across ten tasks. The red line consistently encloses or significantly extends beyond the blue line, indicating superior performance for "AgentFlow" in all categories. * **AgentFlow (w/o Flow-GRPO) (Blue Line):** * MedQA: 76.0 * GPQA: 37.0 * GameOf24: 31.0 * AMC23: 47.4 * AIME24: 16.7 * GAIA: 17.2 * Musique: 25.3 * HotpotQA: 51.3 * 2Wiki: 60.0 * Bamboogle: 58.4 * **AgentFlow (Red Line):** * MedQA: 80.0 * GPQA: 47.0 * GameOf24: 53.0 * AMC23: 61.5 * AIME24: 40.0 * GAIA: 33.1 * Musique: 33.1 * HotpotQA: 57.0 * 2Wiki: 71.2 * Bamboogle: 69.6 * **Aggregated Improvements:** * **Science:** +7.0% (associated with MedQA, GPQA, Bamboogle) * **Search:** +10.1% (associated with 2Wiki, HotpotQA) * **Agentic:** +15.9% (associated with Musique, GAIA) * **Math:** +19.8% (associated with AIME24, AMC23, GameOf24) **Bar Charts (Right Side)** Each bar chart shows the "Accuracy (%)" for different models on a specific task. The red bar, representing "AgentFlow (7B)", consistently shows the highest performance in all six tasks. 1. **2Wiki (Search)** * Trend: AgentFlow (7B) is significantly higher than all other models. * Qwen-2.5-7B (light gray): 23.0 * GPT-4o (~200B) (dark gray): 49.5 * ToRL (7B) (light blue): 38.2 * Search-R1 (7B) (medium blue): 47.6 * AutoGen (7B) (light purple): 44.0 * AgentFlow (7B) (red): 77.2 2. **HotpotQA (Search)** * Trend: AgentFlow (7B) is the highest, with GPT-4o and AutoGen as the next best performers. * Qwen-2.5-7B (light gray): 21.0 * GPT-4o (~200B) (dark gray): 54.0 * ToRL (7B) (light blue): 37.0 * Search-R1 (7B) (medium blue): 43.5 * AutoGen (7B) (light purple): 50.0 * AgentFlow (7B) (red): 57.0 3. **GAIA (Agentic)** * Trend: AgentFlow (7B) shows a substantial lead over all other models, which perform significantly lower. * Qwen-2.5-7B (light gray): 3.2 * GPT-4o (~200B) (dark gray): 17.3 * ToRL (7B) (light blue): 19.1 * Search-R1 (7B) (medium blue): 17.3 * ReSearch (7B) (dark blue): 6.3 * AutoGen (7B) (light purple): Not present (The dark blue bar for ReSearch (7B) is the 6th bar, with value 6.3. The light purple bar for AutoGen (7B) is not present in this chart.) * AgentFlow (7B) (red): 33.1 4. **AIME24 (Math)** * Trend: AgentFlow (7B) is significantly superior, with ToRL (7B) being the second-best but still far behind. * Qwen-2.5-7B (light gray): 6.7 * GPT-4o (~200B) (dark gray): 13.3 * TIR (7B) (light green): 10.0 * ToRL (7B) (light blue): 20.0 * AutoGen (7B) (light purple): 13.3 * AgentFlow (7B) (red): 40.0 5. **GameOf24 (Math)** * Trend: AgentFlow (7B) performs substantially better than other models, which are clustered in a lower range. * Qwen-2.5-7B (light gray): 33.0 * GPT-4o (~200B) (dark gray): 32.0 * TIR (7B) (light green): 33.0 * ToRL (7B) (light blue): 31.0 * AutoGen (7B) (light purple): 24.0 * AgentFlow (7B) (red): 53.0 6. **GPQA (Science)** * Trend: AgentFlow (7B) achieves the highest accuracy, followed by TIR (7B) and Search-R1 (7B). * Qwen-2.5-7B (light gray): 34.0 * GPT-4o (~200B) (dark gray): 31.0 * TIR (7B) (light green): 42.0 * ToRL (7B) (light blue): 35.0 * Search-R1 (7B) (medium blue): 42.0 * AgentFlow (7B) (red): 47.0 ### Key Observations * **Consistent Superiority of AgentFlow:** In both the radar chart and all six bar charts, "AgentFlow" (red line/bar) consistently outperforms its baseline "AgentFlow (w/o Flow-GRPO)" and all other compared models. * **Impact of Flow-GRPO:** The radar chart clearly demonstrates that the "Flow-GRPO" component significantly boosts AgentFlow's performance across all ten tasks, as the red line (with Flow-GRPO) is always outside the blue line (without Flow-GRPO). * **Significant Gains in Math and Agentic Tasks:** The aggregated percentage improvements highlight that "AgentFlow" achieves its largest gains over the baseline in Math tasks (+19.8%) and Agentic tasks (+15.9%). * **Strong Performance in Search Tasks:** AgentFlow shows substantial leads in 2Wiki (77.2%) and HotpotQA (57.0%) compared to other models, with a +10.1% aggregated improvement in Search tasks. * **Varied Competitor Performance:** While AgentFlow is consistently best, the performance of other models varies. GPT-4o (~200B) often performs well among the non-AgentFlow models, particularly in search tasks (e.g., 49.5% in 2Wiki, 54.0% in HotpotQA). * **Large Gaps in Challenging Tasks:** In tasks like GAIA (Agentic) and AIME24 (Math), AgentFlow's lead is particularly pronounced, suggesting its approach is highly effective for these more complex problem types. For instance, in GAIA, AgentFlow scores 33.1% while the next best is ToRL (7B) at 19.1%. ### Interpretation The data strongly suggests that "AgentFlow" is a highly effective system, and its "Flow-GRPO" component is critical for its superior performance. The radar chart provides a holistic view, emphasizing the broad applicability and consistent improvement across diverse task categories, including Science, Search, Agentic, and Math. The aggregated percentage increases further quantify these improvements, highlighting the most impactful areas. The bar charts provide crucial validation by comparing "AgentFlow (7B)" against a range of other established and competitive models. The consistent top performance of "AgentFlow (7B)" across all six detailed tasks (2Wiki, HotpotQA, GAIA, AIME24, GameOf24, GPQA) indicates its robustness and generalizability. The significant performance gaps, especially in tasks like GAIA and AIME24, imply that "AgentFlow" might possess unique capabilities or a more effective strategy for handling the complexities inherent in agentic and mathematical reasoning tasks. The fact that "AgentFlow (7B)" often doubles or triples the accuracy of other 7B models (e.g., in AIME24) and even outperforms much larger models like GPT-4o (~200B) in several instances (e.g., 2Wiki, GAIA, AIME24, GameOf24), underscores its efficiency and effectiveness. Overall, the image serves as compelling evidence for the efficacy of the "AgentFlow" system, particularly when augmented with "Flow-GRPO," positioning it as a leading solution for a wide array of challenging AI benchmarks. The data implies that the "Flow-GRPO" mechanism likely contributes to better task understanding, planning, or execution, leading to these substantial performance gains.

\n ## Radar Chart & Bar Charts: AgentFlow Performance Comparison ### Overview The image presents a comparison of the AgentFlow model's performance against a baseline (AgentFlow w/o Flow-GRPO) across several benchmarks. The comparison is visualized using a radar chart for an overview and a series of bar charts for detailed performance on individual benchmarks. The benchmarks cover diverse areas like question answering (2Wiki, HotpotQA, GAIA, GPQA), reasoning (Math, GameOf24, AMC23, AIME24), and science (MedQA, Musique, Bamboogle). ### Components/Axes * **Radar Chart:** * **Axes:** MedQA, Science, GPQA, GameOf24, Math, AMC23, AIME24, GAIA, 2Wiki, HotpotQA, Musique, Bamboogle. These represent the different benchmarks. * **Scale:** 0 to 80 (approximately). * **Lines:** * AgentFlow (w/o Flow-GRPO) - Blue line * AgentFlow - Red line * **Legend:** Located in the top-left corner. * **Bar Charts:** * **X-axis:** Model names (Qwen-2.5-7B, GPT-4o (~200B), Search-R1 (7B), ReSearch (7B), AutoGen (7B), AgentFlow (7B)). * **Y-axis:** Accuracy (%) - Scale from 0 to 80 (approximately). * **Charts:** 2Wiki (Search), HotpotQA (Search), GAIA (Agentic), AIME24 (Math), GameOf24 (Math), GPQA (Science). * **Legend:** Color-coded bars representing each model. ### Detailed Analysis or Content Details **Radar Chart Analysis:** The radar chart displays the performance of AgentFlow with and without Flow-GRPO across 12 benchmarks. The red line represents AgentFlow *with* Flow-GRPO, and the blue line represents AgentFlow *without* Flow-GRPO. * **MedQA:** AgentFlow (w/o Flow-GRPO): ~80.0%, AgentFlow: ~76.0% * **Science:** AgentFlow (w/o Flow-GRPO): ~76.0%, AgentFlow: ~69.6% * **GPQA:** AgentFlow (w/o Flow-GRPO): ~47.0%, AgentFlow: ~37.0% * **GameOf24:** AgentFlow (w/o Flow-GRPO): ~53.0%, AgentFlow: ~47.4% * **Math:** AgentFlow (w/o Flow-GRPO): ~61.5%, AgentFlow: ~40.0% * **AMC23:** AgentFlow (w/o Flow-GRPO): ~61.5%, AgentFlow: ~31.0% * **AIME24:** AgentFlow (w/o Flow-GRPO): ~17.2%, AgentFlow: ~16.7% * **GAIA:** AgentFlow (w/o Flow-GRPO): ~58.4%, AgentFlow: ~33.1% * **2Wiki:** AgentFlow (w/o Flow-GRPO): ~71.2%, AgentFlow: ~60.0% * **HotpotQA:** AgentFlow (w/o Flow-GRPO): ~51.3%, AgentFlow: ~57.0% * **Musique:** AgentFlow (w/o Flow-GRPO): ~25.3%, AgentFlow: ~19.2% * **Bamboogle:** AgentFlow (w/o Flow-GRPO): ~69.6%, AgentFlow: ~60.0% The chart also indicates overall performance improvements: +7.0% (GPQA), +19.8% (Math), +15.9% (GAIA), +10.1% (2Wiki). **Bar Chart Analysis:** * **2Wiki (Search):** Qwen-2.5-7B: ~49.5%, GPT-4o (~200B): ~72.2%, Search-R1 (7B): ~38.2%, AutoGen (7B): ~44.0%, ReSearch (7B): ~21.0%, AgentFlow (7B): ~23.3% * **HotpotQA (Search):** Qwen-2.5-7B: ~54.0%, GPT-4o (~200B): ~43.5%, Search-R1 (7B): ~37.0%, AutoGen (7B): ~30.0%, ReSearch (7B): ~3.2%, AgentFlow (7B): ~6.3% * **GAIA (Agentic):** Qwen-2.5-7B: ~50.0%, GPT-4o (~200B): ~33.1%, Search-R1 (7B): ~17.3%, AutoGen (7B): ~19.1%, ReSearch (7B): ~6.3%, AgentFlow (7B): ~17.3% * **AIME24 (Math):** Qwen-2.5-7B: ~40.0%, GPT-4o (~200B): ~13.3%, Search-R1 (7B): ~10.0%, AutoGen (7B): ~20.0%, ReSearch (7B): ~6.7%, AgentFlow (7B): ~10.0% * **GameOf24 (Math):** Qwen-2.5-7B: ~53.0%, GPT-4o (~200B): ~31.0%, Search-R1 (7B): ~33.0%, AutoGen (7B): ~30.0%, ReSearch (7B): ~24.0%, AgentFlow (7B): ~33.0% * **GPQA (Science):** Qwen-2.5-7B: ~42.0%, GPT-4o (~200B): ~35.0%, Search-R1 (7B): ~34.0%, AutoGen (7B): ~31.0%, ReSearch (7B): ~47.0%, AgentFlow (7B): ~42.0% ### Key Observations * AgentFlow consistently performs better than the baseline (AgentFlow w/o Flow-GRPO) across all benchmarks in the radar chart. * The largest performance gains with AgentFlow are observed in GPQA and Math. * GPT-4o (~200B) generally achieves the highest accuracy across most benchmarks in the bar charts. * AgentFlow (7B) generally performs lower than GPT-4o (~200B), Qwen-2.5-7B, and Search-R1 (7B) in the bar charts. * ReSearch (7B) shows very low performance in HotpotQA and AIME24. ### Interpretation The data suggests that the Flow-GRPO component significantly improves the performance of the AgentFlow model across a diverse set of tasks. The radar chart provides a holistic view of these improvements, while the bar charts offer a more granular comparison against other models. The consistent outperformance of GPT-4o highlights the current state-of-the-art in large language models. The relatively low performance of AgentFlow (7B) compared to larger models suggests that scaling model size remains a crucial factor in achieving high accuracy. The significant drop in ReSearch (7B)'s performance on HotpotQA and AIME24 could indicate a specific weakness in that model's architecture or training data for those tasks. The combination of radar and bar charts provides a comprehensive assessment of AgentFlow's capabilities and areas for potential improvement.

## Composite Performance Analysis: AgentFlow vs. Baseline Models ### Overview The image is a composite technical figure comparing the performance of an AI system named "AgentFlow" against several baseline models across a variety of reasoning and knowledge-intensive tasks. It consists of two main sections: a radar chart on the left and a grid of six bar charts on the right. The overall theme is a performance benchmark, highlighting AgentFlow's improvements in specific domains. ### Components/Axes **1. Left Section: Radar Chart** * **Title/Legend:** Located at the top. Two series are plotted: * `AgentFlow (w/o Flow-GRPO)` - Blue line with circular markers. * `AgentFlow` - Red line with circular markers. * **Axes (Tasks):** The chart has 10 radial axes, each representing a different benchmark task. Labels are placed around the perimeter: * `MedQA` (Top) * `Bamboogle` * `2Wiki` * `HotpotQA` * `Musique` * `GAIA` (Bottom Right) * `AIME24` * `AMC23` * `GameOf24` * `GPQA` (Left) * **Highlighted Improvements:** Three text boxes with percentage gains are placed around the chart: * Top-Left: `+7.0% Science` (near MedQA/GPQA) * Top-Right: `+10.1% Search` (near 2Wiki/HotpotQA) * Bottom-Left: `+19.8% Math` (near GameOf24/AMC23/AIME24) * Bottom-Right: `+15.9% Agentic` (near GAIA/Musique) **2. Right Section: Bar Chart Grid (6 Charts)** * **Common Legend:** Located below the top row of charts. It defines the color coding for all bar charts: * Light Gray: `Qwen-2.5-7B` * Dark Gray: `GPT-4o (~200B)` * Light Blue: `Search-R1 (7B)` * Medium Blue: `ReSearch (7B)` * Light Green: `TIR (7B)` * Teal: `ToRL (7B)` * Purple: `AutoGen (7B)` * **Red: `AgentFlow (7B)`** (This is the primary subject of comparison). * **Individual Charts:** Each chart has a title indicating the task and category, and a Y-axis labeled `Accuracy (%)`. ### Detailed Analysis **A. Radar Chart Data Points (Approximate Values)** The red line (AgentFlow) generally encloses the blue line (AgentFlow w/o Flow-GRPO), indicating superior performance across all tasks. * **MedQA:** AgentFlow ~80.0, w/o Flow-GRPO ~76.0 * **Bamboogle:** AgentFlow ~69.6, w/o Flow-GRPO ~58.4 * **2Wiki:** AgentFlow ~71.2, w/o Flow-GRPO ~60.0 * **HotpotQA:** AgentFlow ~57.0, w/o Flow-GRPO ~51.3 * **Musique:** AgentFlow ~25.3, w/o Flow-GRPO ~19.2 * **GAIA:** AgentFlow ~33.1, w/o Flow-GRPO ~17.2 * **AIME24:** AgentFlow ~40.0, w/o Flow-GRPO ~16.7 * **AMC23:** AgentFlow ~61.5, w/o Flow-GRPO ~47.4 * **GameOf24:** AgentFlow ~53.0, w/o Flow-GRPO ~31.0 * **GPQA:** AgentFlow ~47.0, w/o Flow-GRPO ~37.0 **B. Bar Chart Data Points (Accuracy %)** * **Top-Left: 2Wiki (Search)** * Qwen-2.5-7B: 23.0 * GPT-4o: 49.5 * Search-R1: 38.2 * ReSearch: 47.6 * **AgentFlow: 77.2** (Highest by a large margin) * **Top-Center: HotpotQA (Search)** * Qwen-2.5-7B: 21.0 * GPT-4o: 54.0 * Search-R1: 37.0 * ReSearch: 43.5 * AutoGen: 50.0 * **AgentFlow: 57.0** (Highest) * **Top-Right: GAIA (Agentic)** * Qwen-2.5-7B: 3.2 * GPT-4o: 17.3 * Search-R1: 19.1 * ReSearch: 17.3 * AutoGen: 6.3 * **AgentFlow: 33.1** (Highest, more than double the next best) * **Bottom-Left: AIME24 (Math)** * Qwen-2.5-7B: 6.7 * GPT-4o: 13.3 * TIR: 10.0 * ToRL: 20.0 * AutoGen: 13.3 * **AgentFlow: 40.0** (Highest, double the next best) * **Bottom-Center: GameOf24 (Math)** * Qwen-2.5-7B: 33.0 * GPT-4o: 32.0 * TIR: 33.0 * ToRL: 31.0 * AutoGen: 24.0 * **AgentFlow: 53.0** (Highest, significant lead) * **Bottom-Right: GPQA (Science)** * Qwen-2.5-7B: 34.0 * GPT-4o: 31.0 * TIR: 42.0 * ToRL: 35.0 * AutoGen: 42.0 * **AgentFlow: 47.0** (Highest) ### Key Observations 1. **Consistent Superiority:** AgentFlow (red bar/line) achieves the highest accuracy in every single task presented across both the radar and bar charts. 2. **Domain-Specific Strengths:** The most dramatic performance gaps are in the **Math** (AIME24, GameOf24) and **Agentic** (GAIA) categories, where AgentFlow often doubles or more than doubles the score of the next-best model. 3. **Search Task Dominance:** In search-oriented tasks (2Wiki, HotpotQA), AgentFlow shows a clear lead, though the margin is slightly smaller than in math/agentic tasks. 4. **Impact of Flow-GRPO:** The radar chart demonstrates that the full AgentFlow system (red) consistently outperforms its ablated version without Flow-GRPO (blue), confirming the contribution of this component. 5. **Model Scale Context:** The bar charts compare the 7B-parameter AgentFlow primarily against other 7B models and the much larger GPT-4o (~200B). AgentFlow outperforms both categories. ### Interpretation This composite figure serves as a strong empirical argument for the effectiveness of the AgentFlow architecture. The data suggests that AgentFlow is not just marginally better but represents a significant step forward, particularly in tasks requiring **multi-step reasoning (Math)** and **autonomous tool/agent use (Agentic tasks like GAIA)**. Its consistent lead over both similarly sized models and a much larger frontier model (GPT-4o) indicates architectural efficiencies rather than brute-force scaling. The radar chart's "improvement bubbles" (+7.0% Science, +10.1% Search, etc.) frame the narrative, guiding the viewer to see the gains as domain-specific breakthroughs. The isolation of the "Flow-GRPO" component in the radar chart provides an ablation study, suggesting this specific technique is a key driver of the overall performance gain. **Language Note:** The image contains English text exclusively for labels, legends, and data. The category labels "Science", "Search", "Math", and "Agentic" are in English.

## Radar Chart: AgentFlow Performance Comparison (w/o Flow-GRPO vs. AgentFlow) ### Overview A radar chart comparing two configurations of AgentFlow across multiple datasets. The blue line represents "AgentFlow (w/o Flow-GRPO)" and the red line represents "AgentFlow". The chart includes labeled axes for datasets and percentage-based performance metrics. ### Components/Axes - **Axes**: - Labeled with datasets: Bambooogle, 2Wiki, HotpotQA, Musiq, GameOf24, AMC23, AIME24, GAIA. - Radial scale: 0% to 80% (approximate). - **Legend**: - Blue: AgentFlow (w/o Flow-GRPO) - Red: AgentFlow - Positioned at the top-left of the radar chart. ### Detailed Analysis - **Blue Line (AgentFlow w/o Flow-GRPO)**: - Values: 69.6 (Bambooogle), 71.2 (2Wiki), 57.0 (HotpotQA), 31.0 (Musiq), 47.4 (GameOf24), 61.5 (AMC23), 40.0 (AIME24), 51.3 (GAIA). - Trend: Peaks at Bambooogle (69.6%) and declines toward Musiq (31.0%). - **Red Line (AgentFlow)**: - Values: 80.0 (Bambooogle), 77.2 (2Wiki), 57.0 (HotpotQA), 25.3 (Musiq), 61.5 (AMC23), 40.0 (AIME24), 51.3 (GAIA). - Trend: Consistently higher than the blue line across all datasets, with a notable drop at Musiq (25.3%). ### Key Observations - **Performance Gains**: - +7.0% improvement in Science (MedQA). - +10.1% improvement in Search (2Wiki). - +19.8% improvement in Math (AIME24). - +15.9% improvement in Agentic tasks (GAIA). - **Outliers**: - Musiq dataset shows the largest drop for both configurations (blue: 31.0% → red: 25.3%). ### Interpretation The radar chart demonstrates that AgentFlow with Flow-GRPO (red line) outperforms the baseline AgentFlow (blue line) across all datasets, with the most significant gains in Math (+19.8%) and Search (+10.1%). The Musiq dataset is an outlier, where both configurations underperform, suggesting potential domain-specific limitations. The annotations highlight AgentFlow's adaptability across diverse tasks, particularly in computational and agentic reasoning. --- ## Bar Charts: Model Performance Across Tasks ### Components/Axes - **Tasks**: - 2Wiki (Search), HotpotQA (Search), GAIA (Agentic), AIME24 (Math), GameOf24 (Math), GPQA (Science). - **Models**: - Qwen-2.5-7B (gray), GPT-4o (~200B) (dark gray), Search-R1 (7B) (light blue), ReSearch (7B) (blue), ToRL (7B) (teal), AutoGen (7B) (purple), AgentFlow (7B) (red). - **Legend**: Positioned at the bottom-right of the bar charts. ### Detailed Analysis - **2Wiki (Search)**: - AgentFlow (7B): 77.2% (highest). - Qwen-2.5-7B: 49.5%. - GPT-4o: 38.2%. - **HotpotQA (Search)**: - AgentFlow (7B): 57.0%. - GPT-4o: 54.0%. - Search-R1: 37.0%. - **GAIA (Agentic)**: - AgentFlow (7B): 33.1%. - ReSearch: 17.3%. - AutoGen: 6.3%. - **AIME24 (Math)**: - AgentFlow (7B): 40.0%. - ToRL: 20.0%. - Qwen-2.5-7B: 6.7%. - **GameOf24 (Math)**: - AgentFlow (7B): 53.0%. - GPT-4o: 33.0%. - AutoGen: 24.0%. - **GPQA (Science)**: - AgentFlow (7B): 47.0%. - ReSearch: 42.0%. - ToRL: 35.0%. ### Key Observations - **AgentFlow Dominance**: - AgentFlow (7B) consistently achieves the highest accuracy across all tasks. - Notable gaps: - 2Wiki: 77.2% (AgentFlow) vs. 49.5% (Qwen-2.5-7B). - AIME24: 40.0% (AgentFlow) vs. 6.7% (Qwen-2.5-7B). - **Model Specialization**: - GPT-4o excels in Search tasks (54.0% in HotpotQA). - ReSearch and ToRL perform moderately in Search and Science tasks. ### Interpretation AgentFlow (7B) demonstrates superior performance across diverse tasks, particularly in Search and Math, where it outperforms larger models like GPT-4o. The GAIA task highlights AgentFlow's edge in agentic reasoning, while ReSearch and ToRL show promise in Search and Science. The stark contrast in AIME24 (Math) underscores AgentFlow's computational reasoning capabilities. These results suggest that AgentFlow's architecture, possibly enhanced by Flow-GRPO, enables efficient task adaptation and accuracy gains.