Papers by Jiazheng Wang

14 papers
Safety Sidecar: Reflection-Driven Runtime Control for Safer Agents (2026.findings-acl)

Copied to clipboard

Challenge: Existing safety controls fail to provide runtime intervention or cross-architecture portability for autonomous LLM agents.
Approach: They propose a model-agnostic, plug-and-play module to provide arbitrary agent safety control and auditability.
Outcome: The proposed module improves the secure-solution rate by 2.9–11.2 percentage points . it adds only 3.2s to end-to-end latency and a negligible average cost of 5.37 10-4 per scenario .
AgentGym2: Benchmarking Large Language Model Agents in De-Idealized Real-World Environments (2026.acl-long)

Copied to clipboard

Challenge: Existing benchmarks evaluate agents in simplified, idealized settings, relying on pre-packaged tool interfaces, overlooking critical steps, and assume inputs are clean and fully specified.
Approach: They propose a framework that evaluates language agents in simplified, idealized settings . they show that even SOTA systems like Gemini and GPT-5 struggle on AgentGym2 .
Outcome: Experiments on 15 proprietary and open-source models show that even SOTA systems like Gemini and GPT-5 struggle on AgentGym2 .
DevEval: A Manually-Annotated Code Generation Benchmark Aligned with Real-World Code Repositories (2024.findings-acl)

Copied to clipboard

Challenge: Existing benchmarks are poorly aligned with real-world code repositories and are insufficient to evaluate the coding abilities of Large Language Models (LLMs).
Approach: They propose a repository-level benchmark named DevEval to evaluate LLMs' coding abilities in real-world code repositories.
Outcome: The proposed benchmarks show that the LLMs perform better in real-world code repositories than existing benchmarks.
DARM: Distribution-Aware Reward Modeling by Alleviating Biases from Low Preference-Context Dependency Data (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for training reward models are vulnerable to context neglect and degraded accuracy.
Approach: They propose distribution-aware reward modeling that augments the RM objective with a conditional mutual information regularizer that maximizes context and the predicted reward conditioned on the response.
Outcome: The proposed model improves performance in RLHF and improves accuracy in other settings.
Multi-Programming Language Sandbox for LLMs (2025.acl-demo)

Copied to clipboard

Challenge: MPLSandbox is an out-of-the-box multi-programming language sandbox designed to provide unified and comprehensive feedback from compiler and analysis tools for Large Language Models (LLMs).
Approach: They propose a multi-programming language sandbox that provides unified feedback from compilers and analysis tools for Large Language Models.
Outcome: The proposed multi-language sandbox can provide comprehensive feedback from compilers and analysis tools for large language models (LLMs).
LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks (2025.acl-long)

Copied to clipboard

Challenge: Existing benchmarks on longcontext large language models fail to reflect their deep understanding capabilities across diverse tasks.
Approach: They propose a benchmark to assess the ability of long-context large language models to handle long-text problems.
Outcome: The proposed model achieves 50.1% accuracy when directly answering the questions . human experts achieve only 53.7% accuracy under a 15-minute time constraint .
SALT: Step-level Advantage Assignment for Long-horizon Agents via Trajectory Graph (2026.findings-eacl)

Copied to clipboard

Challenge: Large Language Models (LLMs) have demonstrated remarkable capabilities, but their application to complex, multi-step, and long-horizon tasks remains challenging.
Approach: They propose a framework that provides a finer-grained advantage assignment derived solely from outcome rewards.
Outcome: The proposed framework provides a finer-grained advantage assignment, derived solely from outcome rewards.
Enhancing Transformers for Generalizable First-Order Logical Entailment (2025.acl-long)

Copied to clipboard

Challenge: Moreover, transformers have demonstrated proficiency in logical reasoning over natural language.
Approach: They propose a logic-aware architecture that improves the performance in generalizable first-order logical entailment by combining distribution shifts and unseen knowledge.
Outcome: The proposed architecture outperforms methods designed specifically for knowledge graph query answering on a dataset with a large dataset.
AlignMMBench: Evaluating Chinese Multimodal Alignment in Large Vision-Language Models (2025.acl-long)

Copied to clipboard

Challenge: Existing benchmarks focus on basic abilities using nonverbal methods, such as yes-no and multiple-choice questions.
Approach: They propose a benchmark that provides more nuanced evaluations of alignment capabilities for large Vision-Language Models (VLMs) they use a rule-calibrated evaluator that exceeds GPT-4's evaluation ability and a “alignment score” to assess the robustness and stability of models across diverse prompts.
Outcome: The proposed benchmark covers 13 tasks across three categories and includes both single-turn and multi-turn dialogue scenarios.
ABC-Bench: Benchmarking Agentic Backend Coding in Real-World Development (2026.findings-acl)

Copied to clipboard

Challenge: Large Language Models (LLMs) have redefined the role of AI in software engineering . current benchmarks focus on localized code generation, but neglect dynamic, full-process requirements of real-world engineering.
Approach: They propose a benchmark to evaluate agentic backend coding within a realistic, executable workflow.
Outcome: The ABC-Bench benchmark evaluates agentic backend coding within a realistic, executable workflow.
LLaMA-Rider: Spurring Large Language Models to Explore the Open World (2024.findings-naacl)

Copied to clipboard

Challenge: Recent studies have used Large Language Models to help decision-making and planning in environments, but their capacity to acquire environmental knowledge and adapt in an open world remains uncertain.
Approach: They propose an approach to spur LLMs to explore the open world, gather experiences, and learn to improve their task-solving capabilities by using a feedback-revision mechanism.
Outcome: The proposed model enhances the efficiency of the LLM in exploring the open world and improves its ability to accomplish more tasks through fine-tuning with merely 1.3k instances of collected data, showing minimal training costs compared to baseline using reinforcement learning.
VRPO: Rethinking Value Modeling for Robust RL under Noisy Supervision in LLM Post-Training (2026.acl-long)

Copied to clipboard

Challenge: Reinforcement Learning (RL) in real-world environments often suffers from ambiguous or incomplete supervision.
Approach: They propose a framework that enhances value modeling for robust RL in LLM post-training by integrating auxiliary losses guided by entropy and perplexity from a frozen language model and variational information bottleneck.
Outcome: The proposed framework outperforms baselines on multi-turn dialogue, math reasoning, and science QA with rule-based and model-based rewards.
DocFusion: A Unified Framework for Document Parsing Tasks (2025.findings-acl)

Copied to clipboard

Challenge: Existing methods for document parsing often employ multiple models, limiting performance . Existing models often employ discrete tokens, whereas recognition relies on continuous coordinates .
Approach: They propose a Gaussian-Kernel Cross-Entropy Loss (GK-CEL) that unifies detection and recognition by enabling generative frameworks to handle both tasks simultaneously.
Outcome: The proposed model performs competitively across four core document parsing tasks.
A.S.E: A Repository-Level Benchmark for Evaluating Security in AI-Generated Code (2026.findings-acl)

Copied to clipboard

Challenge: Existing security evaluation benchmarks lack relevance to real-world AI programming tasks . current LLMs struggle with secure coding, research shows .
Approach: They propose a repository-level evaluation benchmark to assess security of AI-generated code.
Outcome: The proposed framework mirrors real-world AI programming tasks and offers valuable insights into the state of AI code generation.

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