Selected work

Representative Papers

Post-Training on Office Work Improves Software Engineering: A Behavioral Account of Cross-Domain Transfer

Aug 02, 2026

This work addresses the limited goal-directed execution capability of large language models in long-horizon tasks by introducing a Goal-Directed Execution (GDE) behavioral framework. The authors conduct post-training on the Qwen3.5-122B-A10B model using 363 long-horizon, multi-tool agent tasks from office scenarios, without relying on software engineering data. This approach yields a notable improvement on SWE-Bench Pro, increasing pass@1 by 5.8 percentage points. Experimental results demonstrate significant enhancements across four core GDE capabilities: goal selection, state construction, goal consistency maintenance, and environment validation. Furthermore, the model exhibits effective cross-domain transfer between office and software engineering tasks, confirming that long-horizon post-training can successfully drive the transfer of behavioral mechanisms.

0 citationsRead paper

Cross-Benchmark Generalization in Long-Horizon Agents

Jul 31, 2026

Existing reinforcement learning agents often overfit to idiosyncratic patterns in closed environments and lack verifiable behavioral generalization. This work proposes the first cross-domain, long-horizon, multi-tool post-training framework, built upon the open-source MoE model Qwen3.5-122B-A10B and combining two-stage supervised fine-tuning (SFT) with reinforcement learning (RL). Training is conducted on 363 tasks across 27 categories within the MCP benchmark, strictly isolating external evaluation tasks and reward signals. Experimental results demonstrate that the proposed approach substantially enhances out-of-distribution transfer performance, achieving consistent gains across five external benchmarks—including Toolathlon (+9.6 percentage points) and τ²-Bench (+5.3 pp)—and even improves performance on SWE-Bench Pro and Terminal-Bench 2 despite the absence of software engineering tasks in training. The study further uncovers four consistent cross-scenario behavioral divergence patterns.

0 citationsRead paper

GDP.pdf: Benchmarking Grounded Multimodal Reasoning over Professional PDF Documents

Jul 13, 2026

Current document-centric AI benchmarks often assess individual capabilities in isolation, failing to capture models’ holistic reasoning performance on PDFs in authentic professional contexts. To address this gap, this work introduces a high-quality benchmark comprising question–document pairs authored by experts across ten professional domains, retaining only those samples on which at least two state-of-the-art multimodal models commit substantive errors. The study uniquely focuses on real-world querying scenarios involving professional PDFs and proposes an atomic-criteria–based scoring mechanism alongside a three-tier, eleven-category capability taxonomy. It also emphasizes the model’s ability to abstain when queries lack sufficient support in the document. Evaluation reveals that even the best-performing model passes only 15% of the 100 test cases, with primary failure modes including table misalignment, chart misinterpretation, footnote omission, symbol-counting errors, and mishandling of revised or overlaid text.

0 citationsRead paper

ComplexConstraints and Beyond: Expert Rubrics for RLVR

Jun 08, 2026

This work addresses the limitations of existing large language model (LLM) evaluation methods, which struggle to effectively assess complex, context-dependent instruction following and agent behaviors, often focusing only on superficial constraints. To overcome this, we propose a novel evaluation paradigm grounded in expert-designed rubrics, featuring atomic, intent-aware scoring criteria calibrated via LLM-based judges to enable precise assessment and efficient training on complex tasks. We introduce five principles for high-quality rubric design and, for the first time, leverage expert rubrics simultaneously as both evaluation instruments and reinforcement learning signals. Experiments demonstrate that models trained on our ComplexConstraints dataset exhibit substantial improvements—15.5% and 12.2% gains in instruction-following performance for 4B and 235B parameter models, respectively—and show strong generalization to unseen enterprise-level tasks, with notable improvements on BFCL (+4.5%), Tau2-Bench (+7.4%), and Tool-Decathlon (+6.8%).

0 citationsRead paper

Riemann-Bench: A Benchmark for Moonshot Mathematics

Apr 08, 2026

This study addresses the lack of reliable evaluation of artificial intelligence systems on genuine mathematical research, despite their strong performance on Olympiad-style problems. To bridge this gap, the authors introduce a novel private benchmark comprising 25 expert-level problems, meticulously designed by leading mathematicians and validated through double-blind review, emphasizing deep theoretical understanding over procedural problem-solving tricks. Evaluation employs a programmatic verifier to check unique closed-form solutions and an unbiased statistical estimator based on 100 independent runs per problem, permitting models to leverage coding, search, and open-ended reasoning. Results reveal that all state-of-the-art models achieve success rates below 10%, underscoring a substantial disparity between current AI capabilities and human expertise in research-grade mathematical reasoning.

0 citationsRead paper

Recent publications

Latest Papers

Post-Training on Office Work Improves Software Engineering: A Behavioral Account of Cross-Domain Transfer

Aug 02, 2026

This work addresses the limited goal-directed execution capability of large language models in long-horizon tasks by introducing a Goal-Directed Execution (GDE) behavioral framework. The authors conduct post-training on the Qwen3.5-122B-A10B model using 363 long-horizon, multi-tool agent tasks from office scenarios, without relying on software engineering data. This approach yields a notable improvement on SWE-Bench Pro, increasing pass@1 by 5.8 percentage points. Experimental results demonstrate significant enhancements across four core GDE capabilities: goal selection, state construction, goal consistency maintenance, and environment validation. Furthermore, the model exhibits effective cross-domain transfer between office and software engineering tasks, confirming that long-horizon post-training can successfully drive the transfer of behavioral mechanisms.

0 citationsRead paper

Cross-Benchmark Generalization in Long-Horizon Agents

Jul 31, 2026

Existing reinforcement learning agents often overfit to idiosyncratic patterns in closed environments and lack verifiable behavioral generalization. This work proposes the first cross-domain, long-horizon, multi-tool post-training framework, built upon the open-source MoE model Qwen3.5-122B-A10B and combining two-stage supervised fine-tuning (SFT) with reinforcement learning (RL). Training is conducted on 363 tasks across 27 categories within the MCP benchmark, strictly isolating external evaluation tasks and reward signals. Experimental results demonstrate that the proposed approach substantially enhances out-of-distribution transfer performance, achieving consistent gains across five external benchmarks—including Toolathlon (+9.6 percentage points) and τ²-Bench (+5.3 pp)—and even improves performance on SWE-Bench Pro and Terminal-Bench 2 despite the absence of software engineering tasks in training. The study further uncovers four consistent cross-scenario behavioral divergence patterns.

0 citationsRead paper

GDP.pdf: Benchmarking Grounded Multimodal Reasoning over Professional PDF Documents

Jul 13, 2026

Current document-centric AI benchmarks often assess individual capabilities in isolation, failing to capture models’ holistic reasoning performance on PDFs in authentic professional contexts. To address this gap, this work introduces a high-quality benchmark comprising question–document pairs authored by experts across ten professional domains, retaining only those samples on which at least two state-of-the-art multimodal models commit substantive errors. The study uniquely focuses on real-world querying scenarios involving professional PDFs and proposes an atomic-criteria–based scoring mechanism alongside a three-tier, eleven-category capability taxonomy. It also emphasizes the model’s ability to abstain when queries lack sufficient support in the document. Evaluation reveals that even the best-performing model passes only 15% of the 100 test cases, with primary failure modes including table misalignment, chart misinterpretation, footnote omission, symbol-counting errors, and mishandling of revised or overlaid text.

0 citationsRead paper

ComplexConstraints and Beyond: Expert Rubrics for RLVR

Jun 08, 2026

This work addresses the limitations of existing large language model (LLM) evaluation methods, which struggle to effectively assess complex, context-dependent instruction following and agent behaviors, often focusing only on superficial constraints. To overcome this, we propose a novel evaluation paradigm grounded in expert-designed rubrics, featuring atomic, intent-aware scoring criteria calibrated via LLM-based judges to enable precise assessment and efficient training on complex tasks. We introduce five principles for high-quality rubric design and, for the first time, leverage expert rubrics simultaneously as both evaluation instruments and reinforcement learning signals. Experiments demonstrate that models trained on our ComplexConstraints dataset exhibit substantial improvements—15.5% and 12.2% gains in instruction-following performance for 4B and 235B parameter models, respectively—and show strong generalization to unseen enterprise-level tasks, with notable improvements on BFCL (+4.5%), Tau2-Bench (+7.4%), and Tool-Decathlon (+6.8%).

0 citationsRead paper

Riemann-Bench: A Benchmark for Moonshot Mathematics

Apr 08, 2026

This study addresses the lack of reliable evaluation of artificial intelligence systems on genuine mathematical research, despite their strong performance on Olympiad-style problems. To bridge this gap, the authors introduce a novel private benchmark comprising 25 expert-level problems, meticulously designed by leading mathematicians and validated through double-blind review, emphasizing deep theoretical understanding over procedural problem-solving tricks. Evaluation employs a programmatic verifier to check unique closed-form solutions and an unbiased statistical estimator based on 100 independent runs per problem, permitting models to leverage coding, search, and open-ended reasoning. Results reveal that all state-of-the-art models achieve success rates below 10%, underscoring a substantial disparity between current AI capabilities and human expertise in research-grade mathematical reasoning.

0 citationsRead paper