Papers by Zhijiang Li

16 papers
Too Long, Do Re-weighting for Efficient LLM Reasoning Compression (2026.acl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) have recently achieved remarkable progress on complex reasoning tasks by leveraging extended Chain-of-Thought (CoT) techniques.
Approach: They propose a method that uses Extended Chain-of-Thought (EFT) to reduce the number of output tokens by nearly 40% while maintaining the accuracy of the reasoning.
Outcome: The proposed method reduces the number of output tokens by nearly 40% while maintaining the accuracy of the reasoning.
CtrlA: Adaptive Retrieval-Augmented Generation via Inherent Control (2025.findings-acl)

Copied to clipboard

Challenge: Existing methods focus on detecting LLM’s confidence via statistical uncertainty.
Approach: They propose to use a representation perspective to solve adaptive RAG by enabling dynamic retrieval during generation and enabling retrieval only when the query exceeds LLM's internal knowledge.
Outcome: The proposed framework is superior to existing adaptive RAG methods on a diverse set of tasks.
When Inverse Data Outperforms: Exploring the Pitfalls of Mixed Data in Multi-Stage Fine-Tuning (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing methods for o1-level performance focus on unidirectional supervised fine-tuning (SFT), overlooking the intricate interplay between diverse reasoning patterns.
Approach: They construct a reverse reasoning dataset and examine how it is supervised . they find that naively mixing forward and reverse data during SFT weakens the directional distinction .
Outcome: The proposed model improves accuracy by 1.6%–6.8% over a standard model.
LLM Jailbreak Detection for (Almost) Free! (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing methods for detecting jailbreak prompts entail significant computational costs .
Approach: They propose a free jailbreak detection method which scales logits by temperature to detect jailbreak prompts .
Outcome: The proposed method detects jailbreak prompts with no additional computational costs.
TreeReview: A Dynamic Tree of Questions Framework for Deep and Efficient LLM-based Scientific Peer Review (2025.emnlp-main)

Copied to clipboard

Challenge: Large Language Models (LLMs) have shown significant potential in assisting peer review, but current methods struggle to generate thorough and insightful reviews while maintaining efficiency.
Approach: They propose a framework that models paper review as a hierarchical and bidirectional question-answering process.
Outcome: The proposed framework outperforms baselines on full review generation and actionable feedback comments generation tasks while reducing LLM token usage by up to 80% compared to computationally intensive approaches.
DGPO: Beyond Pairwise Preferences with Directional Consistent Groupwise Optimization (2026.findings-acl)

Copied to clipboard

Challenge: Existing methods for directional consistency alignment of large language models are limited . a recent study suggests reverse supervision as a complement to forward reasoning .
Approach: They propose a framework that aggregates supervision signals at the group level and explicitly models direction-aware alignment through multi-candidate comparisons.
Outcome: The proposed framework achieves 3.2% accuracy improvement across five benchmarks and multiple datasets.
TRIGO: Benchmarking Formal Mathematical Proof Reduction for Generative Language Models (2023.emnlp-main)

Copied to clipboard

Challenge: Automated theorem proving (ATP) benchmarks focus on symbolic inference but rarely involve understanding complex number combination reasoning.
Approach: They propose a benchmark that requires a model to reduce a trigonometric expression with step-by-step proof and evaluates a generative LM’s reasoning ability on formulas and ability to manipulate, group, and factor number terms.
Outcome: The proposed benchmark evaluates a generative LM’s reasoning ability on formulas and ability to manipulate, group, and factor number terms.
Reinforcement Learning on Pre-Training Data (2026.acl-long)

Copied to clipboard

Challenge: Recent progress in large language models is driven by scaling of training compute through pre-training with nexttoken prediction (NTP) or post-training (RL) Pre-training using NTP enables models to acquire extensive knowledge and skills from general data, but it suffers from data inefficiency and catastrophic forgetting in continual learning settings.
Approach: They propose to scale training compute through pre-training with next-token prediction (NTP) or post-training by scaling reinforcement learning (RL) to improve learning from general data.
Outcome: Experiments on multiple benchmarks and models show that the proposed approach improves continual pre-training and provides a strong foundation for post-training on Qwen3-8B-Base.
Reimagining Safety Alignment with An Image (2025.emnlp-main)

Copied to clipboard

Challenge: Existing approaches to large language models face inefficiency, fragility, or architectural constraints, resulting in inefficient performance and heightened over-refusal in cross-modal tasks.
Approach: They propose an optimization-driven visual prompt framework that enhances security and reduces over-refusal at the same time.
Outcome: The proposed framework enhances security and reduces over-refusal while maintaining robust safety while reducing unnecessary denials.
FaStFact: Faster, Stronger Long-Form Factuality Evaluations in LLMs (2025.findings-emnlp)

Copied to clipboard

Challenge: Prior evaluation pipelines fail to evaluate factuality of long-form LLMs due to inefficiency and costly human assessment.
Approach: They propose a fast and strong evaluation pipeline that can evaluate factuality of long-form LLMs . they propose 'faStFact' to reduce cost of web searching and inference calling .
Outcome: The proposed evaluation pipeline achieves highest alignment with human evaluation and efficiency among existing baselines.
ProxyQA: An Alternative Framework for Evaluating Long-Form Text Generation with Large Language Models (2024.acl-long)

Copied to clipboard

Challenge: Existing evaluation methods for large language models are labor-intensive and lack efficiency.
Approach: They propose a framework dedicated to assessing long-text generation that includes in-depth human-curated meta-questions spanning various domains . they use a set of proxy-quests with pre-annotated answers to assess the content's quality by incorporating the generated texts as contextual background.
Outcome: The proposed framework assesses the quality of long-text content by matching it with references through human evaluation or automated metrics.
DVD: Dynamic Contrastive Decoding for Knowledge Amplification in Multi-Document Question Answering (2024.emnlp-main)

Copied to clipboard

Challenge: Large language models (LLMs) generate information with hallucinations due to uneven retrieval quality and irrelevant contents.
Approach: They propose a decoding strategy which dynamically amplifies knowledge from selected documents during the generation phase.
Outcome: The proposed method outperforms other decoding strategies on ALCE-ASQA, NQ, TQA and PopQA benchmarks.
You Only Need One Single Token to Refine Safety Alignment (2026.findings-acl)

Copied to clipboard

Challenge: Excessive safety can lead to over-refusal, where models reject harmful-looking yet benign queries, severely limiting utility.
Approach: They propose a lightweight training-based approach that reshapes the distributions of harmful and benign samples within the model’s decision space by using a single-token prefix.
Outcome: The proposed approach can distinguish between harmful and benign samples while keeping the model frozen.
Evaluating Robustness of Generative Search Engine on Adversarial Factoid Questions (2024.findings-acl)

Copied to clipboard

Challenge: Existing large language models (LLMs)-backed generative search engines may not always be accurate.
Approach: They propose to evaluate the robustness of retrieval-augmented generation in a realistic and high-risk setting where adversaries have only black-box system access.
Outcome: The proposed model exhibits higher susceptibility to factual errors compared to LLMs without retrieval.
Beyond End-to-End: Dynamic Chain Optimization for Private LLM Adaptation on the Edge (2026.acl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) are revolutionizing mobile intelligence, but their implementation on mobile devices is severely bottlenecked by the prohibitive resource demands of LLMs.
Approach: They propose a paradigm that forgoes end-to-end updates in favor of a sequential, layer-by-layer manner.
Outcome: Extensive experiments on multiple benchmarks demonstrate the superiority of ChainFed over existing methods, boosting average accuracy by up to 46.46%.
Spotlight and Shadow: Attention-Guided Dual-Anchor Introspective Decoding for MLLM Hallucination Mitigation (2026.findings-acl)

Copied to clipboard

Challenge: Multimodal Large Language Models (MLLMs) excel in tasks ranging from image captioning to complex reasoning.
Approach: They propose a contrastive decoding framework that dynamically calibrates each token generation by mining the model’s internal perceptual discrepancies.
Outcome: The proposed framework mitigates hallucination while enhancing general reasoning capabilities.

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