Selected work

Representative Papers

Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Jul 30, 2026

This work presents the first practical implementation of recursive self-improvement (RSI) in machine learning engineering, introducing a unified learn-and-evolve loop framework. The authors develop OpenMLE, a full-stack executable AI4AI system that integrates a verifiable task environment, operator learning, and long-horizon search modules. They apply execution-driven supervised fine-tuning and reinforcement learning to the Frontis-MA1 (35B) model, transforming it into a meta-evolutionary agent that performs closed-loop optimization through four atomic operations: Draft, Improve, Debug, and Crossover. Leveraging deduplicated training data and an asynchronous experience prior mechanism (OpenMLE-Evo-Max), the system achieves a Medal Average of 71.21% on MLE-Bench Lite—surpassing GPT-5.5+Codex—and attains a 70% Match-SOTA score on NatureBench Lite, demonstrating strong generalization and cross-domain capabilities.

0 citationsRead paper

NatureBench: Can Coding Agents Match the Published SOTA of Nature-Family Papers?

Jun 23, 2026

This study investigates whether AI coding agents can achieve original scientific discoveries beyond mere replication on real-world research problems. To this end, the authors introduce NatureBench, the first benchmark comprising 90 executable, cross-disciplinary tasks derived from papers published in the Nature family of journals, along with NatureGym—a standardized evaluation framework featuring containerized environments, automated build pipelines, and a strict prohibition on web search to address fragmentation and unreliable assessment in scientific task evaluation. Experimental results show that even the strongest current agents surpass the original papers’ state-of-the-art performance on only 17.8% of tasks, with successes largely attributable to method transfer rather than genuine scientific innovation. Primary failure modes include incorrect method selection and insufficient computational resources.

0 citationsRead paper

EnterpriseClawBench: Benchmarking Agents from Real Workplace Sessions

Jun 22, 2026

This study addresses the lack of realistic benchmarks for evaluating AI agents in authentic enterprise work environments, which hinders comprehensive assessment of their capabilities in complex office settings. The authors introduce the first structured evaluation suite derived from real-world enterprise agent conversations, comprising 852 tasks—each specifying a prompt, role category, skill subcategory, hard constraints, and semantic scoring criteria. They further propose a multidimensional evaluation protocol encompassing model-framework combinations, artifact delivery, visual quality, cost, runtime efficiency, and skill transferability. Experimental results reveal that even the best-performing configuration (Codex + GPT-5.5) achieves only a score of 0.663, underscoring the significant limitations of current agents on enterprise-grade tasks and affirming the necessity and value of developing multidimensional, real-scenario evaluation benchmarks.

0 citationsRead paper

Post-Trained MoE Can Skip Half Experts via Self-Distillation

May 18, 2026

This work addresses the challenge of efficiently converting static Mixture-of-Experts (MoE) models into dynamically sparsely activated ones without re-pretraining, thereby reducing inference costs. The authors propose Zero-Expert Self-Distillation Adaptation (ZEDA), a framework that enables post-training dynamic adaptation of static MoE models without fine-tuning. ZEDA integrates parameter-free zero-output experts, a two-stage self-distillation process, group-level load-balancing loss, and a dynamic expert-skipping mechanism. Evaluated on Qwen3-30B-A3B and GLM-4.7-Flash, ZEDA skips over 50% of expert FLOPs on average with minimal accuracy degradation, outperforming the strongest dynamic MoE baseline by 6.1 and 4.0 points, respectively, while achieving approximately 1.2× end-to-end inference speedup.

0 citationsRead paper

TIR-Agent: Training an Explorative and Efficient Agent for Image Restoration

Mar 29, 2026

Existing image restoration agents rely on untrained heuristic scheduling and exhaustive tool invocation, resulting in suboptimal restoration pathways and high computational overhead. This work proposes TIR-Agent, a trainable image restoration agent that learns efficient task scheduling and tool composition strategies through a two-stage training paradigm combining supervised fine-tuning and reinforcement learning. The approach innovatively incorporates stochastic perturbations to enhance exploration and introduces a multidimensional adaptive reward mechanism that dynamically fuses image quality metrics to mitigate reward hacking. Experimental results demonstrate that TIR-Agent outperforms twelve baseline models across both in-domain and out-of-domain degradation scenarios, achieving over 2.5× faster inference and significantly reducing redundant tool calls.

0 citationsRead paper

Recent publications

Latest Papers

Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Jul 30, 2026

This work presents the first practical implementation of recursive self-improvement (RSI) in machine learning engineering, introducing a unified learn-and-evolve loop framework. The authors develop OpenMLE, a full-stack executable AI4AI system that integrates a verifiable task environment, operator learning, and long-horizon search modules. They apply execution-driven supervised fine-tuning and reinforcement learning to the Frontis-MA1 (35B) model, transforming it into a meta-evolutionary agent that performs closed-loop optimization through four atomic operations: Draft, Improve, Debug, and Crossover. Leveraging deduplicated training data and an asynchronous experience prior mechanism (OpenMLE-Evo-Max), the system achieves a Medal Average of 71.21% on MLE-Bench Lite—surpassing GPT-5.5+Codex—and attains a 70% Match-SOTA score on NatureBench Lite, demonstrating strong generalization and cross-domain capabilities.

0 citationsRead paper

NatureBench: Can Coding Agents Match the Published SOTA of Nature-Family Papers?

Jun 23, 2026

This study investigates whether AI coding agents can achieve original scientific discoveries beyond mere replication on real-world research problems. To this end, the authors introduce NatureBench, the first benchmark comprising 90 executable, cross-disciplinary tasks derived from papers published in the Nature family of journals, along with NatureGym—a standardized evaluation framework featuring containerized environments, automated build pipelines, and a strict prohibition on web search to address fragmentation and unreliable assessment in scientific task evaluation. Experimental results show that even the strongest current agents surpass the original papers’ state-of-the-art performance on only 17.8% of tasks, with successes largely attributable to method transfer rather than genuine scientific innovation. Primary failure modes include incorrect method selection and insufficient computational resources.

0 citationsRead paper

EnterpriseClawBench: Benchmarking Agents from Real Workplace Sessions

Jun 22, 2026

This study addresses the lack of realistic benchmarks for evaluating AI agents in authentic enterprise work environments, which hinders comprehensive assessment of their capabilities in complex office settings. The authors introduce the first structured evaluation suite derived from real-world enterprise agent conversations, comprising 852 tasks—each specifying a prompt, role category, skill subcategory, hard constraints, and semantic scoring criteria. They further propose a multidimensional evaluation protocol encompassing model-framework combinations, artifact delivery, visual quality, cost, runtime efficiency, and skill transferability. Experimental results reveal that even the best-performing configuration (Codex + GPT-5.5) achieves only a score of 0.663, underscoring the significant limitations of current agents on enterprise-grade tasks and affirming the necessity and value of developing multidimensional, real-scenario evaluation benchmarks.

0 citationsRead paper

Post-Trained MoE Can Skip Half Experts via Self-Distillation

May 18, 2026

This work addresses the challenge of efficiently converting static Mixture-of-Experts (MoE) models into dynamically sparsely activated ones without re-pretraining, thereby reducing inference costs. The authors propose Zero-Expert Self-Distillation Adaptation (ZEDA), a framework that enables post-training dynamic adaptation of static MoE models without fine-tuning. ZEDA integrates parameter-free zero-output experts, a two-stage self-distillation process, group-level load-balancing loss, and a dynamic expert-skipping mechanism. Evaluated on Qwen3-30B-A3B and GLM-4.7-Flash, ZEDA skips over 50% of expert FLOPs on average with minimal accuracy degradation, outperforming the strongest dynamic MoE baseline by 6.1 and 4.0 points, respectively, while achieving approximately 1.2× end-to-end inference speedup.

0 citationsRead paper

TIR-Agent: Training an Explorative and Efficient Agent for Image Restoration

Mar 29, 2026

Existing image restoration agents rely on untrained heuristic scheduling and exhaustive tool invocation, resulting in suboptimal restoration pathways and high computational overhead. This work proposes TIR-Agent, a trainable image restoration agent that learns efficient task scheduling and tool composition strategies through a two-stage training paradigm combining supervised fine-tuning and reinforcement learning. The approach innovatively incorporates stochastic perturbations to enhance exploration and introduces a multidimensional adaptive reward mechanism that dynamically fuses image quality metrics to mitigate reward hacking. Experimental results demonstrate that TIR-Agent outperforms twelve baseline models across both in-domain and out-of-domain degradation scenarios, achieving over 2.5× faster inference and significantly reducing redundant tool calls.

0 citationsRead paper