Papers by Jiaheng Zhang

35 papers
Chinese SafetyQA: A Safety Short-form Factuality Benchmark for Large Language Models (2025.acl-long)

Copied to clipboard

Challenge: Large language models have created significant safety concerns . factuality ability is crucial in determining whether they can be deployed and applied safely and compliantly within specific regions.
Approach: They propose a benchmark to evaluate the factuality of large language models in China . they evaluate the models' ability to provide accurate and reliable information .
Outcome: The proposed benchmark evaluates the factuality abilities of existing LLMs and compares them to LLM abilities.
E2-LLM: Efficient and Extreme Length Extension of Large Language Models (2024.findings-acl)

Copied to clipboard

Challenge: Existing techniques for extending context capabilities in LLMs require additional training procedures and access to datasets with long context (e.g., sequences of 32K tokens).
Approach: They propose a solution to extend context capabilities in Large Language Models by training a single process over a sequence of 4K tokens.
Outcome: The proposed solution significantly reduces the cost of continual-pretraining or fine-tuning over short sequences and improves robustness to diverse relative positions.
USB: A COMPREHENSIVE AND UNIFIED SAFETY EVALUATION BENCHMARK FOR MULTIMODAL LARGE LANGUAGE MODELS (2026.acl-long)

Copied to clipboard

Challenge: Existing safety benchmarks fail to provide reliable assessments due to limited risk coverage, insufficient scale and the oversight of complex modality combinations.
Approach: They propose a framework that covers 61 risk categories across four modality interactions to address this gap.
Outcome: The proposed framework covers 61 risk categories across four distinct modality interactions.
ReviewRL: Towards Automated Scientific Review with RL (2025.emnlp-main)

Copied to clipboard

Challenge: Existing automated review systems struggle with factual accuracy, rating consistency, and analytical depth.
Approach: They propose a framework for generating comprehensive and factually grounded scientific paper reviews using supervised fine-tuning and reinforcement learning.
Outcome: The proposed framework outperforms existing methods on ICLR 2025 papers.
Skeleton-Guided-Translation: A Benchmarking Framework for Code Repository Translation with Fine-Grained Quality Evaluation (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing code translation benchmarks focus on individual functions, overlooking repository-level challenges like intermodule coherence and dependency management.
Approach: They propose a framework for benchmarking Java-to-C# translation at the repository level . it uses a translation framework guided by skeletons and fine-grained quality evaluation .
Outcome: The proposed framework improves Java-to-C# translation quality at the repository level.
DI-BENCH: Benchmarking Large Language Models on Dependency Inference with Testable Repositories at Scale (2025.findings-acl)

Copied to clipboard

Challenge: Existing studies highlight that dependency-related issues cause over 40% of observed runtime errors on the generated repository.
Approach: They propose a large-scale benchmark and evaluation framework specifically designed to assess LLMs’ capability on dependency inference.
Outcome: The proposed model achieves only a 48% execution pass rate on Python, indicating room for improvement.
VidCapBench: A Comprehensive Benchmark of Video Captioning for Controllable Text-to-Video Generation (2025.findings-acl)

Copied to clipboard

Challenge: Existing studies have not identified a link between video caption evaluation and T2V generation.
Approach: They propose a video caption evaluation scheme specifically designed for T2V generation that integrates video annotation with caption evaluation.
Outcome: The proposed system is agnostic to any particular caption format and can be used for training.
MT-Video-Bench: A Holistic Video Understanding Benchmark for Evaluating Multimodal LLMs in Multi-Turn Dialogues (2026.findings-acl)

Copied to clipboard

Challenge: Existing evaluation benchmarks for Multimodal Large Language Models (MLLMs) focus on single-turn question answering, overlooking the complexity of multi-turn dialogues in real-world scenarios.
Approach: They propose a video understanding benchmark for MLLMs in multi-turn dialogues that assesses six core competencies that focus on perceptivity and interactivity.
Outcome: The MT-Video-Bench evaluates 1,000 multi-turn dialogues from diverse domains and reveals significant performance discrepancies and limitations in handling multi-turned video dialogues.
ReLook: Vision-Grounded RL with a Multimodal LLM Critic for Agentic Web Coding (2026.acl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) excel at algorithmic code generation, but front-end development is lacking in visual fidelity and interaction.
Approach: They propose an agentic, vision-grounded reinforcement learning framework that closes a loop by invoking a multimodal LLM as a tool.
Outcome: The proposed framework outperforms baselines in front-end code generation.
VIDA: A Visual Intent-driven Design Assistant for Proactive Multimodal Clarification (2026.findings-acl)

Copied to clipboard

Challenge: Existing vision-language models fail to provide accurate and complete answers to user requests . a new strategy-aware design assistant is developed to help designers create proactive, visually grounded, and strategically prioritized clarification questions.
Approach: They propose a visual intent-driven design assistant to generate proactive, visually grounded, and strategically prioritized clarification questions.
Outcome: The proposed assistant improves the strategic alignment score by 20.59% over baselines and restores visual grounding capabilities lost during fine-tuning.
Beyond Hard Masks: Progressive Token Evolution for Diffusion Language Models (2026.acl-long)

Copied to clipboard

Challenge: Existing Diffusion Language Models rely on hard binary masking and discrete token assignments, which hinder the revision of early decisions.
Approach: They propose a diffusion-based language modeling approach that replaces hard binary masks with evolving soft token distributions.
Outcome: The proposed approach outperforms existing DLMs on multiple benchmarks.
Efficient Self-Evaluation for Diffusion Language Models via Sequence Regeneration (2026.acl-long)

Copied to clipboard

Challenge: Non-sequential and bidirectional nature of diffusion large language models makes direct likelihood-based self-evaluation challenging.
Approach: They propose a self-evaluation confidence quantification method for diffusion large language models that quantifies confidence by computing the probability of regenerating tokens in the entire generated sequence, given the full context.
Outcome: The proposed method is correlated with semantic coherence and answer accuracy.
MdEval: Massively Multilingual Code Debugging (2026.findings-acl)

Copied to clipboard

Challenge: Existing benchmarks primarily focus on Python and are limited in terms of language diversity.
Approach: They propose a multilingual debugging benchmark that includes 3.9K test samples of 20 programming languages and introduces the debug instruction corpora MdEval-Instruct by injecting bugs into the correct multilingual queries and solutions.
Outcome: The proposed benchmark includes 3.9K test samples of 20 programming languages and covers the automated program repair task, bug localization task, and bug identification task.
CriticLean: Critic-Guided Reinforcement Learning for Mathematical Formalization (2026.acl-long)

Copied to clipboard

Challenge: Existing approaches to formalizing mathematical statements face limitations in accuracy, especially in the context of complex, highlevel problems that involve sophisticated mathematical reasoning.
Approach: They propose a CriticLean framework that elevates the role of the critic from a passive validator to an active learning component and introduce a benchmark to measure models’ ability to distinguish semantically correct from incorrect formalizations.
Outcome: The proposed framework outperforms open- and closed-source benchmarks and shows that it significantly outperformed existing models.
Safety in Large Reasoning Models: A Survey (2025.findings-emnlp)

Copied to clipboard

Challenge: Large Reasoning Models (LRMs) have a high level of advanced reasoning capabilities, but they are vulnerable and vulnerable.
Approach: This paper presents the first comprehensive survey of Large Reasoning Models . it explores the new safety risks, attacks, and defense strategies specific to LRMs based on reasoning .
Outcome: The proposed study examines the safety and security risks of large reasoning models.
Can MLLMs Understand the Deep Implication Behind Chinese Images? (2025.acl-long)

Copied to clipboard

Challenge: MLLMs perform poorly on traditional culture images, indicating limitations in understanding high-level semantics and lacking a deep knowledge base of Chinese traditional culture.
Approach: They propose to use Chinese images to assess MLLMs' higher-order perception and understanding of Chinese visual content.
Outcome: The proposed model incorporates images that represent Chinese traditional culture, such as famous Chinese traditional paintings, to ensure the authenticity of the Chinese context.
LVP-M3: Language-aware Visual Prompt for Multilingual Multimodal Machine Translation (2022.emnlp-main)

Copied to clipboard

Challenge: Recent advances struggle to train a separate model for each language pair, which is costly and unaffordable when the number of languages increases in the real world.
Approach: They propose to train different MMT models to support translations between different languages.
Outcome: The proposed model is able to handle the above issues by providing a shared semantic space for multiple languages.
MIO: A Foundation Model on Multimodal Tokens (2025.emnlp-main)

Copied to clipboard

Challenge: Existing models lack multimodal understanding capabilities, resulting in closed-source model that does not support multimodal interleaved sequences.
Approach: They propose a foundation model built on multimodal tokens capable of understanding and generating speech, text, images, and videos in an end-to-end, autoregressive manner.
Outcome: The proposed model is able to understand speech, text, images, and videos in an end-to-end, autoregressive manner.
Unsupervised Sentence Textual Similarity with Compositional Phrase Semantics (2022.coling-1)

Copied to clipboard

Challenge: Sentence Textual Similarity (STS) is a classic task that can be applied to downstream NLP applications such as text generation and retrieval.
Approach: They propose a light-weighted Expectation-Correction (EC) formulation for STS computation.
Outcome: The proposed approach is more efficient and scalable than previous approaches.
Thinking-Based Non-Thinking: Solving the Reward Hacking Problem in Training Hybrid Reasoning Models via Reinforcement Learning (2026.acl-long)

Copied to clipboard

Challenge: Existing work on large reasoning models (LRMs) focuses on using reinforcement learning (RL) to train hybrid reasoning models that automatically decide whether to engage in thinking or not based on the complexity of the query.
Approach: They propose to use reinforcement learning to train hybrid reasoning models that automatically decide whether to engage in thinking or not based on the complexity of the query.
Outcome: The proposed model reduces token usage by around 50%$ compared to DeepSeek-R1-Distill-Qwen-1.5B/7B and DeepScaleR-1.5b, while significantly improving accuracy.
M3TQA: Massively Multilingual Multitask Table Question Answering (2026.findings-acl)

Copied to clipboard

Challenge: Existing multilingual table benchmarks suffer from geolinguistic imbalance - overrepresenting certain languages and lacking sufficient scale for rigorous cross-lingual analysis.
Approach: They propose a framework for massively multilingual table question answering that includes tables expanded to 97 languages from Chinese and English sources.
Outcome: Experiments on state-of-the-art LLMs show that synthetically generated training data significantly boosts performance, especially for low-resource languages.
COIG-P: A High-Quality and Large-Scale Chinese Preference Dataset for Alignment with Human Values (2026.findings-eacl)

Copied to clipboard

Challenge: Existing Chinese preference datasets suffer from limited scale, restricted domain coverage, and insufficiently rigorous data validation.
Approach: They propose an LLM-based data annotation pipeline with no human intervention to annotate Chinese preference datasets.
Outcome: The proposed pipeline outperforms existing Chinese preference datasets on AlignBench and Chinese Reward Benchmark.
DREAM: Disentangling Risks to Enhance Safety Alignment in Multimodal Large Language Models (2025.naacl-long)

Copied to clipboard

Challenge: Multimodal Large Language Models (MLLMs) pose unique safety challenges due to their integration of visual and textual data.
Approach: They propose a method to disentangle risks through step-by-step reasoning within multimodal inputs.
Outcome: The proposed approach improves safety alignment in MLLMs by fine-tuning and iterative Reinforcement Learning from AI feedback.
RoleLLM: Benchmarking, Eliciting, and Enhancing Role-Playing Abilities of Large Language Models (2024.findings-acl)

Copied to clipboard

Challenge: Large Language Models (LLMs) have paved the way for complex tasks such as role-playing.
Approach: They propose a framework to benchmark, elicit, and enhance role-playing abilities in Large Language Models.
Outcome: The proposed framework improves role-playing abilities with 168,093 samples.
OpenCoder: The Open Cookbook for Top-Tier Code Large Language Models (2025.acl-long)

Copied to clipboard

Challenge: Code LLMs lack reproducible data pipelines and training protocols for reproducible advancements in code intelligence.
Approach: They propose a top-tier code LLM that releases model weights and inference code . reproducible data pipelines, rigorous experimental ablation results and training protocols are included .
Outcome: The proposed model achieves comparable performance to leading models and serves as an "open cookbook" reproducible training data, rigorous experimental ablation results, and detailed training protocols are also included in the model.
MMRA: A Benchmark for Evaluating Multi-Granularity and Multi-Image Relational Association Capabilities in Large Visual Language Models (2026.findings-eacl)

Copied to clipboard

Challenge: Current multimodal benchmarks focus on facts within individual images, but neglect associative relations among multiple images.
Approach: They propose a multi-image relational association task and a MMRA benchmark to evaluate LVLMs.
Outcome: The proposed benchmarks show that entity-level multi-image perception tasks pose greater challenges than image-level tasks.
Towards Real-world Scenario: Imbalanced New Intent Discovery (2024.acl-long)

Copied to clipboard

Challenge: Existing studies focus on detecting known and previously undefined categories of user intent . skewed and long-tailed distributions often encountered in open-world scenarios .
Approach: They propose to use imbalanced new intent discovery task to identify familiar and novel intent categories within long-tailed distributions.
Outcome: The proposed model outperforms the existing benchmark on three datasets to simulate the real-world long-tail distributions.
OAgents: An Empirical Study of Building Effective Agents (2025.findings-emnlp)

Copied to clipboard

Challenge: a recent study shows that agent research practices are far from standard, rigorous . lack of a standard evaluation protocol makes previous works not reproducible, authors say .
Approach: They conduct an empirical study on the GAIA benchmark to investigate agent design choices . they find that lack of a standard evaluation protocol makes previous works not reproducible .
Outcome: The proposed framework achieves state-of-the-art performance among open-source projects.
ConceptMath: A Bilingual Concept-wise Benchmark for Measuring Mathematical Reasoning of Large Language Models (2024.findings-acl)

Copied to clipboard

Challenge: ConceptMath evaluates concept-wise mathematical reasoning of Large Language Models (LLMs) Existing benchmarks that evaluate general mathematical reasoning with an average accuracy fail to probe the fine-grained failure modes of mathematical reasoning on specific datasets.
Approach: They introduce a bilingual, fine-grained benchmark that evaluates concept-wise mathematical reasoning of Large Language Models.
Outcome: The proposed benchmarks evaluate concept-wise mathematical reasoning of Large Language Models with concept-based accuracies.
Efficient Test-Time Scaling of Multi-Step Reasoning by Probing Internal States of Large Language Models (2026.acl-long)

Copied to clipboard

Challenge: Existing verification approaches, such as Process Reward Models, are computationally expensive and limited to specific domains.
Approach: They propose a transformer-based probe that uses internal states of frozen LLMs to estimate credibility of reasoning steps during generation.
Outcome: The proposed probes match or exceed PRMs that are up to 810 larger.
M2RC-EVAL: Massively Multilingual Repository-level Code Completion Evaluation (2025.acl-long)

Copied to clipboard

Challenge: Existing repository-level code completion benchmarks focus on a limited number of languages . existing benchmarks report overall average scores of different languages ignoring fine-grained abilities .
Approach: They propose to use repository-level code completion benchmarks to evaluate general code intelligence abilities across languages for existing code Large Language Models.
Outcome: The proposed benchmarks improve the code completion abilities of existing LLMs by using two types of annotations on the parsed syntax tree.
CoTJudger: A Graph-Driven Framework for Automatic Evaluation of Chain-of-Thought Efficiency and Redundancy in LRMs (2026.findings-acl)

Copied to clipboard

Challenge: Existing evaluations emphasize final accuracy or coarse token counts, and lack automated tools to separate essential logic from structural redundancy.
Approach: They propose a graph-driven framework that quantifies reasoning efficiency by converting free-form CoTs into directed dependency graphs and extracting the Shortest Effective Path needed to reach a correct solution.
Outcome: Evaluating 21 LRMs, the proposed framework quantifies reasoning efficiency by converting free-form CoTs into directed dependency graphs and extracting the Shortest Effective Path (SEP) needed to reach a correct solution.
LIME: Less Is More for MLLM Evaluation (2025.findings-acl)

Copied to clipboard

Challenge: Existing MLLM benchmarks and unified evaluation frameworks cannot accurately and efficiently reflect the ability of MLMLs.
Approach: They propose a semi-automated benchmark curated using a pipeline that filters out uninformative samples and eliminates answer leakage by focusing on tasks that require image-based understanding.
Outcome: The proposed benchmark reduces the number of samples by 76% and evaluation time by 77% while it can more effectively distinguish different models’ abilities.
When Agents Look the Same: Quantifying Distillation-Induced Similarity in Tool-Use Behaviors (2026.acl-long)

Copied to clipboard

Challenge: Existing metrics fail to distinguish mandatory behaviors required for task success from non-mandatory patterns that reflect a model’s autonomous preferences.
Approach: They propose to use response pattern similarity and action graph similarity to isolate non-mandatory behaviors from mandatory behaviors.
Outcome: Evaluating 18 models from 8 providers on -Bench and 2-Bench against Claude Sonnet 4.5, the authors find that within-family model pairs score 5.9 pp higher in response pattern similarity and action graph similarity .
Can Large Language Models Detect Errors in Long Chain-of-Thought Reasoning? (2025.acl-long)

Copied to clipboard

Challenge: Recent advances in o1-like models have generated long Chain-of-Thought reasoning steps to improve the reasoning abilities of existing Large Language Models (LLMs).
Approach: They propose a DeltaBench to analyze the quality and effectiveness of o1-like models and measure their ability to detect errors in long COT reasoning.
Outcome: The proposed model can detect errors in long COT reasoning.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations