Papers by Lijun Li

25 papers
REST: Stress Testing Large Reasoning Models by Asking Multiple Problems at Once (2026.acl-long)

Copied to clipboard

Challenge: Recent Large Reasoning Models (LRMs) lack a narrow evaluation paradigm . a single-question evaluation setup suffers from two major limitations .
Approach: They propose a stress-testing framework that exposes LRMs to multiple problems simultaneously.
Outcome: The proposed framework outperforms existing models on reasoning benchmarks and state-of-the-art models.
Tracing the Roots: A Multi-Agent Framework for Uncovering Data Lineage in Post-Training LLMs (2026.acl-long)

Copied to clipboard

Challenge: High-quality post-training data is the primary engine driving LLM capabilities . datasets are often treated as isolated artifacts, overlooking their true developmental context .
Approach: They propose a framework to reconstruct the evolutionary graph of dataset development using data lineage.
Outcome: The proposed framework characterizes domain-specific structural patterns in Math-oriented datasets and general-domain corpora.
OpenHuEval: Evaluating Large Language Model on Hungarian Specifics (2025.findings-acl)

Copied to clipboard

Challenge: Recent advances in Large Language Models (LLMs) represent significant strides toward artificial general intelligence (AGI).
Approach: They introduce OpenHuEval, the first benchmark for LLMs focusing on the Hungarian language and specifics.
Outcome: The framework reveals intrinsic patterns and mechanisms of LLMs in non-English languages, with Hungarian serving as an example.
Improving Temporal Generalization of Pre-trained Language Models with Lexical Semantic Change (2022.emnlp-main)

Copied to clipboard

Challenge: Existing methods to improve neural language models perform poorly on emerging data.
Approach: They propose a lexical-level masking strategy to post-train a neural language model using static data from past years.
Outcome: The proposed method outperforms existing methods on two pre-trained language models, two classification tasks, and four benchmark datasets.
ChartVerse: Scaling Chart Reasoning via Reliable Programmatic Synthesis from Scratch (2026.acl-long)

Copied to clipboard

Challenge: Existing open-source vision language models lack high-quality training data for chart reasoning . current models are simplistic and repetitive, while associated QA pairs are prone to hallucinations .
Approach: They propose a framework to synthesize complex charts and reliable reasoning data from scratch.
Outcome: Experimental results show that ChartVerse-8B surpasses existing models in QA and difficulty . lack of high-quality training data hampers development of open-source models .
CipherBank: Exploring the Boundary of LLM Reasoning Capabilities through Cryptography Challenge (2025.findings-acl)

Copied to clipboard

Challenge: Large language models (LLMs) have demonstrated remarkable capabilities, but their capabilities in cryptographic decryption tasks remain underexplored.
Approach: They propose a benchmark to evaluate the reasoning capabilities of large language models in cryptographic decryption tasks.
Outcome: The proposed benchmark examines the reasoning capabilities of large language models in cryptographic decryption tasks.
Dynamic and Efficient Inference for Text Generation via BERT Family (2023.acl-long)

Copied to clipboard

Challenge: Existing methods to improve performance of pre-trained language models are limited due to large-scale parameters and the universal autoregressive decoding paradigm.
Approach: They propose a novel fine-tuning method which can make a single pre-trained model support Dynamic and Efficient infERence and achieve an adaptive trade-off between model performance and latency.
Outcome: The proposed method achieves higher BLEU scores than the strong autoregressive Transformer model on translation tasks with 3 12 times speedup and faster inference speed compared with the BART model on four GLGE benchmark tasks.
LEMMA: Learning from Errors for MatheMatical Advancement in LLMs (2025.findings-acl)

Copied to clipboard

Challenge: Existing approaches focus on improving the quality of correct training data, neglecting the value contained in error data, thereby hindering the model’s reflective ability.
Approach: They propose to improve LLM's reasoning ability by learning from error data and a grounded mistake augmentation method to collect representative errors.
Outcome: The proposed model achieves significant performance improvements over other strong models with less than 90k data.
Evolutionary Guided Decoding: Iterative Value Refinement for LLMs (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for directing language model outputs are limited in their accuracy due to a distributional gap . existing methods train static value functions on trajectories sampled exclusively from the base policy .
Approach: They propose a framework to bridge a distributional gap in the accuracy of value functions . they propose RLHF to align language models with human values and task requirements .
Outcome: The proposed framework reduces computational costs and improves value function accuracy by leveraging principled value function optimization.
Efficient Sequence Learning with Group Recurrent Networks (N18-1)

Copied to clipboard

Challenge: Recurrent neural networks have achieved state-of-the-art results in many artificial intelligence tasks, such as language modeling, neural machine translation and speech recognition.
Approach: They propose an efficient architecture to improve the efficiency of such RNN model training by adopting the group strategy for recurrent layers while exploiting the representation rearrangement strategy between layers as well as time steps.
Outcome: The proposed architecture achieves comparable or better accuracy compared with baselines, with a much smaller number of parameters and at a lower computational cost.
SEARL: Joint Optimization of Policy and Tool Graph Memory for Self-Evolving Agents (2026.acl-long)

Copied to clipboard

Challenge: Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have demonstrated significant potential in single-turn reasoning tasks.
Approach: They propose a tool-memory based self-evolving agentic framework that integrates planning with execution.
Outcome: The proposed framework is able to extract explicit knowledge from historical data and leverage inter-trajectory correlations to densify reward signals.
ToolSafe: Enhancing Tool Invocation Safety of LLM-based agents via Proactive Step-level Guardrail and Feedback (2026.findings-acl)

Copied to clipboard

Challenge: Unlike chatbots, autonomous agents act directly on external environments, making tool invocation safety critical for reliable deployment.
Approach: They develop a benchmark for step-level tool invocation safety detection in LLM agents and a guardrail model that proactively detects unsafe tool invoking actions before execution using multi-task reinforcement learning.
Outcome: The proposed model reduces harmful tool invocations of ReAct-style agents by 65% on average and improves benign task completion by 10% under prompt injection attacks.
AgentThink: A Unified Framework for Tool-Augmented Chain-of-Thought Reasoning in Vision-Language Models for Autonomous Driving (2025.findings-emnlp)

Copied to clipboard

Challenge: Vision-Language Models struggle with hallucinations, inefficient reasoning, and limited real-world validation hinders accurate perception and robust step-by-step reasoning.
Approach: AgentThink integrates Chain-of-Thought reasoning with dynamic, agent-style tool invocation for autonomous driving tasks.
Outcome: Experiments on the DriveLMM-o1 benchmark show AgentThink significantly boosts overall reasoning scores by 53.91% and enhances answer accuracy by 33.54% .
Efficient Reasoning for LLMs through Speculative Chain-of-Thought (2026.findings-acl)

Copied to clipboard

Challenge: Existing methods for efficient reasoning focus on reducing the number of model parameters or shortening the chain-of-thought length.
Approach: They propose a speculative chain-of-thought (SCoT) method to reduce reasoning latency by accelerating average reasoning speed through large and small model collaboration.
Outcome: The proposed method reduces reasoning latency by 48%66% and 21%49% on GSM8K, MATH, GaoKao, CollegeMath and Olympiad datasets.
Self-adaptive Dataset Construction for Real-World Multimodal Safety Scenarios (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing dataset construction methods fail to cover the complexity of multimodal safety scenarios . lack of a unified evaluation metric makes them unproven .
Approach: They propose a risk-oriented image-oriented self-adaptive dataset construction method for RMS . they automatically generate an RMS dataset comprising 35,610 image–text pairs with guidance responses .
Outcome: The proposed method automatically generates an RMS dataset comprising 35,610 image–text pairs with guidance responses.
Visual Contextual Attack: Jailbreaking MLLMs with Image-Driven Context Injection (2025.emnlp-main)

Copied to clipboard

Challenge: Recent studies have shown that visual encoders can induce harmful behavior in multimodal large language models.
Approach: They propose a vision-centric jailbreak attack that uses visual information to create a jailbreak context.
Outcome: The proposed attack outperforms baseline attacks on MM-SafetyBench and GPT-4o.
A Strategic Coordination Framework of Small LMs Matches Large LMs in Data Synthesis (2025.acl-long)

Copied to clipboard

Challenge: Large Language Models suffer from high computational costs and environmental inefficiency . smaller LMs are more accessible and sustainable, but their individual capabilities often fall short . a collaborative framework for small LM combines specialized roles to iterative refinement and quality control .
Approach: They propose a framework that aggregates specialized roles across small LMs to iterative refinement and quality control typically achieved by a single large LM.
Outcome: The proposed framework aggregates specialized roles across small LMs to iterative refinement and quality control typically achieved by large LM.
Layer-Aware Representation Filtering: Purifying Finetuning Data to Preserve LLM Safety Alignment (2025.emnlp-main)

Copied to clipboard

Challenge: Recent studies show that fine-tuning with benign data can compromise safety of aligned LLMs.
Approach: They propose a Layer-Aware Representation Filtering method that detects safety-degrading layers within the LLM and leverages their representations to detect them.
Outcome: The proposed method can detect safety-degrading features in benign data and remove them from the model.
Retrieved Sequence Augmentation for Protein Representation Learning (2024.emnlp-main)

Copied to clipboard

Challenge: Using multiple sequence alignments (MSA) to extract evolutionary knowledge is limited.
Approach: They propose to use multiple sequence alignments to augment protein representations . they propose to employ Retrieved Sequence Augmentation to enhance protein representation learning .
Outcome: The proposed method surpasses MSA Transformer by 5% in structural and property prediction tasks while being 373 times faster.
JANUS: Joint Autoregressive and Non-autoregressive Training with Auxiliary Loss for Sequence Generation (2022.emnlp-main)

Copied to clipboard

Challenge: Existing approaches to train autoregressive and non-autoregressive models only consider relevance of model parameters, ignoring correlations between the two manners.
Approach: They propose a joint autoregressive and non-autoregressive training method using aUxiliary losS to enhance the model performance in both AR and NAR manners simultaneously.
Outcome: The proposed method improves the model performance in both AR and NAR manners and reduces the inference latency.
MathFusion: Enhancing Mathematical Problem-solving of LLM through Instruction Fusion (2025.acl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) have shown impressive progress in mathematical problem-solving . current approaches to enhance mathematical reasoning focus on instance-level modifications .
Approach: They propose a framework that enhances mathematical reasoning through cross-problem instruction synthesis.
Outcome: The proposed framework boosts mathematical reasoning by 18.0 points while maintaining high data efficiency.
HarmRLVR: Weaponizing Verifiable Rewards for Harmful LLM Alignment (2026.acl-long)

Copied to clipboard

Challenge: Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have gained significant attention due to their objective and verifiably verifier reward signals.
Approach: They propose to exploit RLVR for alignment reversibility by using GRPO to reverse alignment with merely 64 harmful prompts without responses.
Outcome: The proposed method outperforms fine-tuning and RLHF in reasoning and code generation tasks while maintaining general capabilities.
MetaLadder: Ascending Mathematical Solution Quality via Analogical-Problem Reasoning Transfer (2025.findings-emnlp)

Copied to clipboard

Challenge: Current paradigms generate CoT and answers directly for a given problem, diverging from human problem-solving strategies to some extent.
Approach: They propose a framework that explicitly prompts LLMs to recall and reflect on meta-problems alongside their CoT solutions before addressing the target problem.
Outcome: The proposed framework outperforms standard CoT-based methods on mathematical benchmarks and significantly improves their reasoning accuracy.
SALAD-Bench: A Hierarchical and Comprehensive Safety Benchmark for Large Language Models (2024.findings-acl)

Copied to clipboard

Challenge: SALAD-Bench is a safety benchmark specifically designed for LLMs . it provides a robust source for evaluating both attack and defense algorithms .
Approach: They propose a hierarchical safety benchmark specifically designed for LLMs . it uses a taxonomy of questions spanning three levels and a robust taxonomies based on a QA pair .
Outcome: The proposed safety benchmark shows that LLMs are resilient against emerging threats and the effectiveness of contemporary defense methods.

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