Selected work

Representative Papers

Scale Can't Overcome Pragmatics: The Impact of Reporting Bias on Vision-Language Reasoning

Feb 26, 2026

This study addresses the limited reasoning capabilities of current vision-language models (VLMs), which stem from reporting bias in training data that systematically omits implicit information—such as spatial relations, temporal dynamics, negation, and counting. For the first time, the authors formally integrate pragmatic theories of reporting bias into vision-language learning, constructing a targeted evaluation benchmark to assess multiple models, including OpenCLIP, LLaVA-1.5, and Molmo. Their findings reveal that merely scaling up data volume or incorporating multilingual corpora fails to rectify these reasoning gaps. In contrast, explicitly designing and integrating annotations that surface such implicit information substantially enhances model performance. This work challenges the prevailing “scale-is-all-you-need” paradigm, underscoring the necessity of deliberately curating training data that supports robust multimodal reasoning.

1 citationsRead paper

FrontierFinance: A Challenging Benchmark for Measuring Frontier Intelligence of Finance Agents

Aug 12, 2026

Existing benchmarks for financial agents are limited to isolated stages of the investment process and fail to evaluate complex reasoning capabilities in realistic settings. This work proposes the first open-ended evaluation framework encompassing the entire investment workflow, featuring 220 expert-crafted multidimensional queries and 11,543 traceable scoring rules across six core use cases. The framework introduces standardized tool-call interfaces grounded in public data and employs an LLM-as-a-judge mechanism calibrated with human oversight. Experiments reveal that toolchain design significantly impacts performance: the Samaya system achieves a leading score of 56.0%, while the open-source model Kimi K3 reaches 46.4%—comparable to top proprietary models at 4.5× lower cost. However, screening and macroeconomic analysis remain key bottlenecks, with even the best system scoring only 33% and 39%, respectively. The dataset and scoring code are publicly released.

0 citationsRead paper

BlendFusion -- Scalable Synthetic Data Generation for Diffusion Model Training

Apr 10, 2026

This work addresses model autophagy disorder (MAD)—a training instability in pure diffusion models caused by visual inconsistencies in generated images—by proposing BlendFusion, a novel framework that integrates path-traced rendering with an object-centric camera placement strategy to construct FineBLEND, a high-quality synthetic text-image dataset. BlendFusion incorporates robust image filtering, automated quality assessment, and aligned text-image annotation mechanisms, enabling highly configurable and scalable 3D data generation. Experimental results demonstrate that FineBLEND significantly outperforms existing mainstream text-image datasets in both image fidelity and training efficacy, effectively mitigating MAD and validating the superiority of object-aware sampling strategies.

0 citationsRead paper

OpaqueToolsBench: Learning Nuances of Tool Behavior Through Interaction

Feb 16, 2026

This work addresses the challenges large language models face when using poorly documented and behaviorally opaque tools, where existing automated documentation methods are either costly or ineffective. The authors propose ToolObserver, a novel framework that dynamically refines tool documentation by iteratively analyzing execution feedback from tool invocation traces, combined with a lightweight exploration strategy to enable efficient and low-cost tool behavior learning. To evaluate this approach, they introduce OpaqueToolsBench, a benchmark encompassing three task categories: function calling, interactive chess playing, and long-horizon intelligent search. Experimental results demonstrate that ToolObserver significantly outperforms baseline methods in complex scenarios, reducing total token consumption during test-time tool exploration by 3.5–7.5× while maintaining high performance.

0 citationsRead paper

Lost in the Maze: Overcoming Context Limitations in Long-Horizon Agentic Search

Oct 21, 2025

Long-horizon agent search suffers from context window limitations, constrained tool budgets, and accumulating redundant information—leading to hallucinations and premature termination. To address these challenges, we propose SLIM, a novel framework that decouples retrieval from browsing and introduces a lightweight, periodic trajectory summarization mechanism to reduce contextual overhead while extending effective exploration depth. SLIM comprises three core components: stepwise retrieval, dynamic content summarization, and an automated fine-grained trajectory analysis pipeline, enabling multi-foundational-model consensus verification. Evaluated on the o3 model, SLIM achieves 56% and 31% accuracy on BrowseComp and HLE benchmarks, respectively—surpassing the best open-source baseline by 8 and 4 percentage points. Moreover, it reduces tool invocations by 4–6×, significantly improving efficiency, robustness, and scalability for long-horizon search tasks.

0 citationsRead paper

Recent publications

Latest Papers

FrontierFinance: A Challenging Benchmark for Measuring Frontier Intelligence of Finance Agents

Aug 12, 2026

Existing benchmarks for financial agents are limited to isolated stages of the investment process and fail to evaluate complex reasoning capabilities in realistic settings. This work proposes the first open-ended evaluation framework encompassing the entire investment workflow, featuring 220 expert-crafted multidimensional queries and 11,543 traceable scoring rules across six core use cases. The framework introduces standardized tool-call interfaces grounded in public data and employs an LLM-as-a-judge mechanism calibrated with human oversight. Experiments reveal that toolchain design significantly impacts performance: the Samaya system achieves a leading score of 56.0%, while the open-source model Kimi K3 reaches 46.4%—comparable to top proprietary models at 4.5× lower cost. However, screening and macroeconomic analysis remain key bottlenecks, with even the best system scoring only 33% and 39%, respectively. The dataset and scoring code are publicly released.

0 citationsRead paper

BlendFusion -- Scalable Synthetic Data Generation for Diffusion Model Training

Apr 10, 2026

This work addresses model autophagy disorder (MAD)—a training instability in pure diffusion models caused by visual inconsistencies in generated images—by proposing BlendFusion, a novel framework that integrates path-traced rendering with an object-centric camera placement strategy to construct FineBLEND, a high-quality synthetic text-image dataset. BlendFusion incorporates robust image filtering, automated quality assessment, and aligned text-image annotation mechanisms, enabling highly configurable and scalable 3D data generation. Experimental results demonstrate that FineBLEND significantly outperforms existing mainstream text-image datasets in both image fidelity and training efficacy, effectively mitigating MAD and validating the superiority of object-aware sampling strategies.

0 citationsRead paper

Scale Can't Overcome Pragmatics: The Impact of Reporting Bias on Vision-Language Reasoning

Feb 26, 2026

This study addresses the limited reasoning capabilities of current vision-language models (VLMs), which stem from reporting bias in training data that systematically omits implicit information—such as spatial relations, temporal dynamics, negation, and counting. For the first time, the authors formally integrate pragmatic theories of reporting bias into vision-language learning, constructing a targeted evaluation benchmark to assess multiple models, including OpenCLIP, LLaVA-1.5, and Molmo. Their findings reveal that merely scaling up data volume or incorporating multilingual corpora fails to rectify these reasoning gaps. In contrast, explicitly designing and integrating annotations that surface such implicit information substantially enhances model performance. This work challenges the prevailing “scale-is-all-you-need” paradigm, underscoring the necessity of deliberately curating training data that supports robust multimodal reasoning.

1 citationsRead paper

OpaqueToolsBench: Learning Nuances of Tool Behavior Through Interaction

Feb 16, 2026

This work addresses the challenges large language models face when using poorly documented and behaviorally opaque tools, where existing automated documentation methods are either costly or ineffective. The authors propose ToolObserver, a novel framework that dynamically refines tool documentation by iteratively analyzing execution feedback from tool invocation traces, combined with a lightweight exploration strategy to enable efficient and low-cost tool behavior learning. To evaluate this approach, they introduce OpaqueToolsBench, a benchmark encompassing three task categories: function calling, interactive chess playing, and long-horizon intelligent search. Experimental results demonstrate that ToolObserver significantly outperforms baseline methods in complex scenarios, reducing total token consumption during test-time tool exploration by 3.5–7.5× while maintaining high performance.

0 citationsRead paper

Lost in the Maze: Overcoming Context Limitations in Long-Horizon Agentic Search

Oct 21, 2025

Long-horizon agent search suffers from context window limitations, constrained tool budgets, and accumulating redundant information—leading to hallucinations and premature termination. To address these challenges, we propose SLIM, a novel framework that decouples retrieval from browsing and introduces a lightweight, periodic trajectory summarization mechanism to reduce contextual overhead while extending effective exploration depth. SLIM comprises three core components: stepwise retrieval, dynamic content summarization, and an automated fine-grained trajectory analysis pipeline, enabling multi-foundational-model consensus verification. Evaluated on the o3 model, SLIM achieves 56% and 31% accuracy on BrowseComp and HLE benchmarks, respectively—surpassing the best open-source baseline by 8 and 4 percentage points. Moreover, it reduces tool invocations by 4–6×, significantly improving efficiency, robustness, and scalability for long-horizon search tasks.

0 citationsRead paper