Institution profile

Zhipu AI

Industry researchasia · cn
Official website
Research library66linked papers
Opportunities0open roles
Selected work

Representative Papers

IF-RewardBench: Benchmarking Judge Models for Instruction-Following Evaluation

Mar 05, 2026

Existing instruction-following meta-evaluation benchmarks suffer from insufficient data coverage and oversimplified evaluation paradigms, limiting their ability to accurately reflect the performance of discriminative models in real-world alignment scenarios. To address this, this work proposes IF-RewardBench, a comprehensive benchmark encompassing diverse instruction types and constraints, which introduces—for the first time—a listwise ranking evaluation paradigm based on multi-response preference graphs. This approach better aligns with practical alignment requirements and significantly enhances the correlation between evaluation outcomes and downstream task performance. Experimental results reveal substantial deficiencies in current discriminative models’ instruction-following capabilities, while demonstrating that IF-RewardBench achieves stronger positive correlation and greater evaluative validity compared to existing benchmarks.

1 citationsRead paper

SWE-Dev: Building Software Engineering Agents with Training and Inference Scaling

Jun 09, 2025

Software engineering (SWE) agents face performance bottlenecks due to scarcity of high-quality training data and insufficient reliable test cases. Method: This paper introduces an open-source large language model agent framework tailored for realistic SWE tasks. Its core innovations are: (1) a robust pipeline for synthesizing *verification-aware* test cases—ensuring functional correctness and behavioral fidelity; and (2) a scalable data construction methodology integrating trajectory distillation, tool-augmented reasoning, and synthetic-test-driven reinforcement learning to generate high-quality agent trajectories. Contribution/Results: Evaluated on the SWE-bench-Verified benchmark, our released models—SWE-Dev 7B and SWE-Dev 32B—achieve success rates of 23.4% and 36.6%, respectively, setting new state-of-the-art results among open-source SWE agents. All code, model weights, and training data are publicly released to foster reproducibility and community advancement.

1 citationsRead paper

GFD-OPD: Guidance-Folded On-Policy Distillation of Diffusion Models Across Scales

Sep 30, 2026

This study addresses the failure of standard online distillation methods from large to small diffusion models, where classifier-free guidance (CFG) amplifies distributional discrepancies between teacher and student. To overcome this challenge, we propose the GFD-OPD framework. We first introduce a novel Fixed-State KL divergence metric that reveals the error accumulation mechanism induced by CFG across cross-scale models. Subsequently, we design a guidance folding strategy that effectively narrows the teacher-student gap and suppresses error propagation, thereby enabling efficient online policy distillation. Extensive experiments demonstrate that our approach significantly improves both training efficiency and generation quality, achieving state-of-the-art performance across all evaluated benchmarks.

0 citationsRead paper

ScopeIF: Improving Scope-Aware Precise Instruction-Following in Large Language Models via Graded Reward Modeling

Sep 26, 2026

This study addresses the imprecise instruction-following behavior of large language models in complex scenarios, which arises from constraint scope omissions and sparse supervision. To tackle this, we propose ScopeIF, a novel framework that introduces the first constraint decomposition schema and constructs the ScopeInstruct dataset. By integrating tool-grounded verification with hierarchical reward reinforcement learning, our approach achieves scope-aware optimization under dense supervision. Experimental results demonstrate that ScopeIF significantly enhances model capability in handling complex scope constraints. Notably, the Qwen3-based implementation outperforms frontier models such as Gemini across multiple evaluation metrics. This work establishes an efficient new paradigm for scope-aware instruction following in large language models.

0 citationsRead paper

From Anomalies to Failures: Constructing Causal Error Graphs for Agentic Trace Diagnosis

Sep 26, 2026

This study addresses the challenges of conflating anomalies with failures and the absence of causal propagation in fault diagnosis over long trajectories of LLM agents. To this end, it proposes the CEG-Agent framework, which first establishes an explicit taxonomy distinguishing anomalies, errors, and failures. It further introduces a unified, typed causal error graph representation and integrates a tool-augmented architecture with a multi-agent adversarial arbitration mechanism to achieve precise fault attribution. Experimental results demonstrate that the proposed method attains state-of-the-art performance on the CEG-Bench benchmark across both semantic and structural evaluations, with its annotations exhibiting strong alignment with expert consensus.

0 citationsRead paper
Recent publications

Latest Papers

GFD-OPD: Guidance-Folded On-Policy Distillation of Diffusion Models Across Scales

Sep 30, 2026

This study addresses the failure of standard online distillation methods from large to small diffusion models, where classifier-free guidance (CFG) amplifies distributional discrepancies between teacher and student. To overcome this challenge, we propose the GFD-OPD framework. We first introduce a novel Fixed-State KL divergence metric that reveals the error accumulation mechanism induced by CFG across cross-scale models. Subsequently, we design a guidance folding strategy that effectively narrows the teacher-student gap and suppresses error propagation, thereby enabling efficient online policy distillation. Extensive experiments demonstrate that our approach significantly improves both training efficiency and generation quality, achieving state-of-the-art performance across all evaluated benchmarks.

0 citationsRead paper

ScopeIF: Improving Scope-Aware Precise Instruction-Following in Large Language Models via Graded Reward Modeling

Sep 26, 2026

This study addresses the imprecise instruction-following behavior of large language models in complex scenarios, which arises from constraint scope omissions and sparse supervision. To tackle this, we propose ScopeIF, a novel framework that introduces the first constraint decomposition schema and constructs the ScopeInstruct dataset. By integrating tool-grounded verification with hierarchical reward reinforcement learning, our approach achieves scope-aware optimization under dense supervision. Experimental results demonstrate that ScopeIF significantly enhances model capability in handling complex scope constraints. Notably, the Qwen3-based implementation outperforms frontier models such as Gemini across multiple evaluation metrics. This work establishes an efficient new paradigm for scope-aware instruction following in large language models.

0 citationsRead paper

From Anomalies to Failures: Constructing Causal Error Graphs for Agentic Trace Diagnosis

Sep 26, 2026

This study addresses the challenges of conflating anomalies with failures and the absence of causal propagation in fault diagnosis over long trajectories of LLM agents. To this end, it proposes the CEG-Agent framework, which first establishes an explicit taxonomy distinguishing anomalies, errors, and failures. It further introduces a unified, typed causal error graph representation and integrates a tool-augmented architecture with a multi-agent adversarial arbitration mechanism to achieve precise fault attribution. Experimental results demonstrate that the proposed method attains state-of-the-art performance on the CEG-Bench benchmark across both semantic and structural evaluations, with its annotations exhibiting strong alignment with expert consensus.

0 citationsRead paper

Enabling a Unified Cross-Domain Representation for Two-Finger Gripper Manipulation via Interaction-Centric Modeling

Sep 25, 2026

This study addresses the challenge of cross-embodiment generalization in imitation learning, which arises from the entanglement of task semantics with hardware-specific visual geometry. To overcome this, we propose an interaction-centric unified framework that generates canonical representations through a parameterized universal gripper abstraction. This representation is integrated with Vision-Language Model (VLM) reasoning, SAM 2.1 segmentation, artificial potential fields, and a Flow-Matching Transformer for hybrid feature-based action prediction. The proposed method uniquely unifies competitive benchmark performance with zero-shot Sim-to-Real transfer under extreme cross-embodiment and cross-view conditions. Extensive evaluations on both simulated environments and real-world heterogeneous robotic platforms validate the framework’s robust cross-domain manipulation capabilities.

0 citationsRead paper