Selected work

Representative Papers

M2G-Eval: Enhancing and Evaluating Multi-granularity Multilingual Code Generation

Dec 27, 2025

Existing code LLM evaluation benchmarks suffer from coarse-grained assessment and limited language coverage, failing to capture fine-grained cross-lingual capability disparities. To address this, we propose M2G-Eval-Coder—the first code generation evaluation framework supporting four granularities (class, function, code block, and line) across 18 programming languages, comprising over 17K training tasks and 1,286 contamination-controlled, human-annotated test samples. We introduce a novel multi-granularity + multilingual co-evaluation paradigm, systematically benchmarking 30 models. Our analysis reveals an ascending difficulty trend from line- to class-level generation, distinct performance patterns between full- and partial-language-support models, and measurable cross-lingual conceptual transferability; we further confirm strong cross-lingual performance correlation. Leveraging supervised fine-tuning (SFT) and GRPO optimization, our approach achieves significant gains over baselines across all granularities.

0 citationsRead paper

Context as a Tool: Context Management for Long-Horizon SWE-Agents

Dec 26, 2025

To address semantic drift, reasoning degradation, and context explosion in long-horizon software engineering agents operating over ultra-large-scale codebases—stemming from uncontrolled context growth—this paper introduces the “Context-as-Tool” (CAT) paradigm, which explicitly models context management as callable, learnable tools. Methodologically: (1) we construct a structured workspace that decouples high-fidelity short-term interactions from compressed long-term memory; (2) we design CAT-GENERATOR, a trajectory-level supervised framework enabling milestone-driven proactive compression; and (3) we develop SWE-Compressor, a context-aware compression model. Evaluated on SWE-Bench-Verified, our approach achieves a 57.6% task success rate—significantly outperforming ReAct baselines and static compression methods—while ensuring robustness and scalability of long-range reasoning under fixed context budgets.

0 citationsRead paper

Universal Reasoning Model

Dec 16, 2025

The performance gains of Universal Transformers (UTs) on complex reasoning tasks—such as ARC-AGI and Sudoku—remain poorly understood, hindering principled architectural design. Method: This work identifies recurrent inductive bias and strong nonlinearity—not structural complexity—as the primary drivers of UT’s reasoning capability. To operationalize these insights, we propose the Universal Reasoning Model (URM): it incorporates lightweight short convolutions for enhanced local pattern modeling, truncated backpropagation through time (BPTT) for improved training stability and long-range reasoning, and position-aware attention with adaptive depth unfolding. Contribution/Results: URM achieves 53.8% and 16.0% pass@1 on ARC-AGI 1 and ARC-AGI 2, respectively—setting new state-of-the-art results at the time. Crucially, this is the first work to systematically isolate, analyze, and empirically validate the roles of key inductive biases in UTs. By decoupling architectural components and linking them to reasoning mechanisms, URM establishes a new paradigm for designing efficient, interpretable, and reasoning-capable neural architectures.

0 citationsRead paper

Scaling Laws for Code: Every Programming Language Matters

Dec 15, 2025

This study addresses the undermodeling of linguistic heterogeneity in multilingual code large language models (LLMs). We propose the first scaling law for multilingual code LLMs that explicitly accounts for language-specific proportions. Based on over 1,000 controlled experiments (equivalent to 336,000 H800 GPU hours), we systematically characterize the nonlinear and non-uniform impact of language composition on model performance: interpreted languages (e.g., Python) benefit disproportionately from scale-up, while syntactically similar language pairs exhibit significant synergistic gains. Building on these insights, we introduce two techniques: (i) cross-lingual token allocation optimization and (ii) parallel code translation pair augmentation. Empirical evaluation shows that, compared to uniform token allocation, our approach substantially improves average performance across languages under fixed compute budgets—yielding pronounced gains for high-value languages (e.g., Python) and reduced resource consumption for fast-saturating ones (e.g., Rust), thereby enhancing overall generalization.

0 citationsRead paper

FATE: A Formal Benchmark Series for Frontier Algebra of Multiple Difficulty Levels

Nov 04, 2025

Existing mathematical benchmarks (e.g., IMO) inadequately assess large language models’ (LLMs’) capabilities in advanced abstract algebraic reasoning. Method: We introduce FATE, a formalized algebraic evaluation benchmark comprising two subsets—FATE-H and FATE-X—each containing 100 problems spanning undergraduate to post-PhD qualifying exam difficulty in abstract and commutative algebra. FATE is the first to systematically formalize high-order algebraic concepts absent from Mathlib. We propose a two-stage evaluation framework that disentangles natural-language reasoning from formalization translation, and conduct systematic evaluations of mainstream LLMs using interactive theorem provers, identifying recurrent formalization error patterns. Results: State-of-the-art models achieve only 3% pass@64 on FATE-H and 0% on FATE-X—substantially lower than their performance on competition mathematics—demonstrating FATE’s high difficulty and its critical value in advancing research on formal mathematical reasoning.

0 citationsRead paper

Recent publications

Latest Papers

M2G-Eval: Enhancing and Evaluating Multi-granularity Multilingual Code Generation

Dec 27, 2025

Existing code LLM evaluation benchmarks suffer from coarse-grained assessment and limited language coverage, failing to capture fine-grained cross-lingual capability disparities. To address this, we propose M2G-Eval-Coder—the first code generation evaluation framework supporting four granularities (class, function, code block, and line) across 18 programming languages, comprising over 17K training tasks and 1,286 contamination-controlled, human-annotated test samples. We introduce a novel multi-granularity + multilingual co-evaluation paradigm, systematically benchmarking 30 models. Our analysis reveals an ascending difficulty trend from line- to class-level generation, distinct performance patterns between full- and partial-language-support models, and measurable cross-lingual conceptual transferability; we further confirm strong cross-lingual performance correlation. Leveraging supervised fine-tuning (SFT) and GRPO optimization, our approach achieves significant gains over baselines across all granularities.

0 citationsRead paper

Context as a Tool: Context Management for Long-Horizon SWE-Agents

Dec 26, 2025

To address semantic drift, reasoning degradation, and context explosion in long-horizon software engineering agents operating over ultra-large-scale codebases—stemming from uncontrolled context growth—this paper introduces the “Context-as-Tool” (CAT) paradigm, which explicitly models context management as callable, learnable tools. Methodologically: (1) we construct a structured workspace that decouples high-fidelity short-term interactions from compressed long-term memory; (2) we design CAT-GENERATOR, a trajectory-level supervised framework enabling milestone-driven proactive compression; and (3) we develop SWE-Compressor, a context-aware compression model. Evaluated on SWE-Bench-Verified, our approach achieves a 57.6% task success rate—significantly outperforming ReAct baselines and static compression methods—while ensuring robustness and scalability of long-range reasoning under fixed context budgets.

0 citationsRead paper

Universal Reasoning Model

Dec 16, 2025

The performance gains of Universal Transformers (UTs) on complex reasoning tasks—such as ARC-AGI and Sudoku—remain poorly understood, hindering principled architectural design. Method: This work identifies recurrent inductive bias and strong nonlinearity—not structural complexity—as the primary drivers of UT’s reasoning capability. To operationalize these insights, we propose the Universal Reasoning Model (URM): it incorporates lightweight short convolutions for enhanced local pattern modeling, truncated backpropagation through time (BPTT) for improved training stability and long-range reasoning, and position-aware attention with adaptive depth unfolding. Contribution/Results: URM achieves 53.8% and 16.0% pass@1 on ARC-AGI 1 and ARC-AGI 2, respectively—setting new state-of-the-art results at the time. Crucially, this is the first work to systematically isolate, analyze, and empirically validate the roles of key inductive biases in UTs. By decoupling architectural components and linking them to reasoning mechanisms, URM establishes a new paradigm for designing efficient, interpretable, and reasoning-capable neural architectures.

0 citationsRead paper

Scaling Laws for Code: Every Programming Language Matters

Dec 15, 2025

This study addresses the undermodeling of linguistic heterogeneity in multilingual code large language models (LLMs). We propose the first scaling law for multilingual code LLMs that explicitly accounts for language-specific proportions. Based on over 1,000 controlled experiments (equivalent to 336,000 H800 GPU hours), we systematically characterize the nonlinear and non-uniform impact of language composition on model performance: interpreted languages (e.g., Python) benefit disproportionately from scale-up, while syntactically similar language pairs exhibit significant synergistic gains. Building on these insights, we introduce two techniques: (i) cross-lingual token allocation optimization and (ii) parallel code translation pair augmentation. Empirical evaluation shows that, compared to uniform token allocation, our approach substantially improves average performance across languages under fixed compute budgets—yielding pronounced gains for high-value languages (e.g., Python) and reduced resource consumption for fast-saturating ones (e.g., Rust), thereby enhancing overall generalization.

0 citationsRead paper

FATE: A Formal Benchmark Series for Frontier Algebra of Multiple Difficulty Levels

Nov 04, 2025

Existing mathematical benchmarks (e.g., IMO) inadequately assess large language models’ (LLMs’) capabilities in advanced abstract algebraic reasoning. Method: We introduce FATE, a formalized algebraic evaluation benchmark comprising two subsets—FATE-H and FATE-X—each containing 100 problems spanning undergraduate to post-PhD qualifying exam difficulty in abstract and commutative algebra. FATE is the first to systematically formalize high-order algebraic concepts absent from Mathlib. We propose a two-stage evaluation framework that disentangles natural-language reasoning from formalization translation, and conduct systematic evaluations of mainstream LLMs using interactive theorem provers, identifying recurrent formalization error patterns. Results: State-of-the-art models achieve only 3% pass@64 on FATE-H and 0% on FATE-X—substantially lower than their performance on competition mathematics—demonstrating FATE’s high difficulty and its critical value in advancing research on formal mathematical reasoning.

0 citationsRead paper