Selected work

Representative Papers

Training Advisors for LLM Agents from Task Outcomes

Oct 07, 2026

This study addresses the lack of effective natural language feedback for large language model (LLM) agents in multi-step tasks by proposing Caddie. This method trains a critic model via reinforcement learning, optimizing end-to-end using only task-level success or failure signals without requiring step-level annotations or reference critiques, thereby providing agents with real-time decision-making guidance. Experiments employing a 4B-parameter critic demonstrate that Caddie achieves zero-shot transferability across models and domains. On the MuSiQue benchmark, it improves the success rates of multiple base models by over 25 percentage points and surpasses the performance of Kimi K3 operating without a critic.

0 citationsRead paper

QuTuner: Feature- and Learning-Guided Optimization Pass Tuning for Quantum Compilers

Jul 05, 2026

Existing quantum compilers explore only a limited optimization pass search space and rely solely on static circuit features, making it difficult to accurately predict optimization outcomes. This work proposes QuTuner, a novel framework that, for the first time, integrates static circuit structural features with dynamic optimization-response embeddings to construct an optimization-aware pass representation. QuTuner employs an offline machine learning model to retrieve and rank candidate optimization sequences, augmented by lightweight online fine-tuning to enable adaptive, multi-objective tuning across different compilers. Experimental results on Qiskit and PyTKET demonstrate that QuTuner reduces optimization metrics by 84.85% and 18.68%, respectively, while cutting tuning time by 73.59% and 64.49%, substantially improving both tuning efficiency and effectiveness.

0 citationsRead paper

SlimSpec: Low-Rank Draft LM-Head for Accelerated Speculative Decoding

May 11, 2026

This work addresses the computational bottleneck in speculative decoding caused by the full-vocabulary LM-head of draft models. The authors propose a low-rank parameterized LM-head architecture that compresses internal representations while preserving support for the complete vocabulary, without requiring vocabulary pruning or substantial modifications to the training pipeline. Built upon low-rank matrix decomposition and integrated into the EAGLE-3 framework, the method is validated across diverse target models and benchmarks. Experimental results demonstrate that, compared to the standard LM-head, the proposed approach achieves a 4–5× speedup in draft token generation and an 8–9% improvement in end-to-end inference throughput, while remaining compatible with various draft model architectures.

0 citationsRead paper

SWE-rebench V2: Language-Agnostic SWE Task Collection at Scale

Feb 27, 2026

This work addresses the scarcity of large-scale, multilingual, and reproducible training data for reinforcement learning in software engineering (SWE) tasks. It proposes the first language-agnostic automated pipeline that harvests executable SWE tasks from real-world code repositories, automatically generates installation and testing procedures, and filters low-quality samples through a combination of large language model evaluation and human verification. For the first time, this approach enables standardized collection of SWE tasks with test validation across 20 programming languages. The released dataset comprises over 32,000 executable tasks spanning more than 3,600 repositories and includes over 120,000 samples with installation instructions and test cases, accompanied by fine-grained metadata and diagnostic analyses—significantly advancing reproducible training and evaluation of multilingual SWE agents.

0 citationsRead paper

LK Losses: Direct Acceptance Rate Optimization for Speculative Decoding

Feb 27, 2026

This work proposes a novel loss function, termed LK loss, which directly optimizes the acceptance rate of speculative decoding—departing from conventional approaches that rely on KL divergence as the training objective. By targeting acceptance rate explicitly, the method eliminates the need for additional computational overhead while significantly improving the efficiency of small-capacity draft models. The approach is architecture-agnostic and demonstrates consistent gains across diverse draft model designs and six large target models ranging from 8B to 685B parameters. Evaluated on general, programming, and mathematical tasks, the proposed method achieves an average increase of 8–10% in accepted token length, thereby substantially enhancing inference speedup without compromising accuracy.

0 citationsRead paper

Recent publications

Latest Papers

Training Advisors for LLM Agents from Task Outcomes

Oct 07, 2026

This study addresses the lack of effective natural language feedback for large language model (LLM) agents in multi-step tasks by proposing Caddie. This method trains a critic model via reinforcement learning, optimizing end-to-end using only task-level success or failure signals without requiring step-level annotations or reference critiques, thereby providing agents with real-time decision-making guidance. Experiments employing a 4B-parameter critic demonstrate that Caddie achieves zero-shot transferability across models and domains. On the MuSiQue benchmark, it improves the success rates of multiple base models by over 25 percentage points and surpasses the performance of Kimi K3 operating without a critic.

0 citationsRead paper

QuTuner: Feature- and Learning-Guided Optimization Pass Tuning for Quantum Compilers

Jul 05, 2026

Existing quantum compilers explore only a limited optimization pass search space and rely solely on static circuit features, making it difficult to accurately predict optimization outcomes. This work proposes QuTuner, a novel framework that, for the first time, integrates static circuit structural features with dynamic optimization-response embeddings to construct an optimization-aware pass representation. QuTuner employs an offline machine learning model to retrieve and rank candidate optimization sequences, augmented by lightweight online fine-tuning to enable adaptive, multi-objective tuning across different compilers. Experimental results on Qiskit and PyTKET demonstrate that QuTuner reduces optimization metrics by 84.85% and 18.68%, respectively, while cutting tuning time by 73.59% and 64.49%, substantially improving both tuning efficiency and effectiveness.

0 citationsRead paper

SlimSpec: Low-Rank Draft LM-Head for Accelerated Speculative Decoding

May 11, 2026

This work addresses the computational bottleneck in speculative decoding caused by the full-vocabulary LM-head of draft models. The authors propose a low-rank parameterized LM-head architecture that compresses internal representations while preserving support for the complete vocabulary, without requiring vocabulary pruning or substantial modifications to the training pipeline. Built upon low-rank matrix decomposition and integrated into the EAGLE-3 framework, the method is validated across diverse target models and benchmarks. Experimental results demonstrate that, compared to the standard LM-head, the proposed approach achieves a 4–5× speedup in draft token generation and an 8–9% improvement in end-to-end inference throughput, while remaining compatible with various draft model architectures.

0 citationsRead paper

SWE-rebench V2: Language-Agnostic SWE Task Collection at Scale

Feb 27, 2026

This work addresses the scarcity of large-scale, multilingual, and reproducible training data for reinforcement learning in software engineering (SWE) tasks. It proposes the first language-agnostic automated pipeline that harvests executable SWE tasks from real-world code repositories, automatically generates installation and testing procedures, and filters low-quality samples through a combination of large language model evaluation and human verification. For the first time, this approach enables standardized collection of SWE tasks with test validation across 20 programming languages. The released dataset comprises over 32,000 executable tasks spanning more than 3,600 repositories and includes over 120,000 samples with installation instructions and test cases, accompanied by fine-grained metadata and diagnostic analyses—significantly advancing reproducible training and evaluation of multilingual SWE agents.

0 citationsRead paper

LK Losses: Direct Acceptance Rate Optimization for Speculative Decoding

Feb 27, 2026

This work proposes a novel loss function, termed LK loss, which directly optimizes the acceptance rate of speculative decoding—departing from conventional approaches that rely on KL divergence as the training objective. By targeting acceptance rate explicitly, the method eliminates the need for additional computational overhead while significantly improving the efficiency of small-capacity draft models. The approach is architecture-agnostic and demonstrates consistent gains across diverse draft model designs and six large target models ranging from 8B to 685B parameters. Evaluated on general, programming, and mathematical tasks, the proposed method achieves an average increase of 8–10% in accepted token length, thereby substantially enhancing inference speedup without compromising accuracy.

0 citationsRead paper