Institution profile

Zhipu AI

Industry researchasia · cn
Official website
Research library62linked 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
Recent publications

Latest Papers

Training Vision-Language-Action Models with Dense Embodied Chain-of-Thought Supervision

Jun 29, 2026

This work addresses the challenge of cross-embodiment transfer for vision–language–action models arising from heterogeneous state and action spaces across robotic platforms. To this end, we propose ZR-0, the first model to incorporate dense Embodied Chain-of-Thought (ECoT) supervision, enabling cognitive alignment across embodiments. ZR-0 features a decoupled dual-stream architecture: a vision–language model generates structured reasoning, while a diffusion Transformer-based action expert produces continuous action chunks; these streams are coupled via constrained cross-attention. Notably, the reasoning stream can be entirely bypassed during deployment without performance degradation. Experiments on LIBERO, RoboTwin 2.0, RoboCasa GR-1, and a real-world xArm platform demonstrate that ZR-0 significantly outperforms existing methods, exhibiting strong cross-embodiment generalization capabilities.

0 citationsRead paper

Video2Code: Generating Interactive Webpages from UI Videos via Action-Aware Revisit

Jun 16, 2026

Existing approaches for translating UI videos into code struggle to accurately capture action boundaries due to sparse temporal representations, leading to misaligned state-action-state transitions and failure to reconstruct authentic user interactions. This work proposes Video2Code, which formulates code generation as an executable state-transition recovery task. It introduces a novel action-aware temporal revisiting mechanism: first identifying key interaction regions at a coarse granularity, then revisiting them with high temporal resolution for fine-grained refinement. By integrating action-aligned video-to-code supervised training with fine-grained temporal cropping, the method substantially enhances the temporal modeling capacity of vision-language models. Experiments demonstrate that Video2Code significantly outperforms existing methods in functional correctness, particularly in complex, multi-step interactive scenarios.

0 citationsRead paper