Papers by Xiaobo Li

21 papers
LearnAlign: Data Selection for LLM Reinforcement Learning with Improved Gradient Alignment (2026.findings-acl)

Copied to clipboard

Challenge: Reinforcement learning with verifiable rewards (RLVR) is a key technique for enhancing LLMs’ reasoning abilities, yet its data inefficiency remains a major bottleneck.
Approach: They propose a gradient-alignment-based method which intelligently selects the learnable and representative training reasoning data for RLVR post-training.
Outcome: Experiments on five reasoning benchmarks show that the proposed method significantly reduces training data requirements while improving performance.
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.
MMA: Cross-Domain Knowledge Integration via Mixture of Multi-Domain Agents (2025.findings-emnlp)

Copied to clipboard

Challenge: achieving synergistic improvements between generalization and domain specialization remains a challenge in pre-training and post-training.
Approach: They propose a test-time cross-domain knowledge integration method that integrates general-purpose and domain-specific models to enhance their performance on complex, domainspecific tasks.
Outcome: The proposed method combines the outputs of general-purpose and domain-specific models to improve their performance on complex, domainspecific tasks.
Adaptive Preference Optimization with Uncertainty-aware Utility Anchor (2025.findings-emnlp)

Copied to clipboard

Challenge: Offline preference optimization methods are efficient for large language models (LLMs) alignment.
Approach: They propose an offline preference optimization framework that estimates uncertainties from preference data . the method enables training even in scenarios where the data is unpaired .
Outcome: The proposed method enables training even in scenarios where the data is unpaired .
MMEvol: Empowering Multimodal Large Language Models with Evol-Instruct (2025.findings-acl)

Copied to clipboard

Challenge: a new framework for image-text instruction data evolution improves MLLM performance . lack of high-quality instruction data remains a major bottleneck in ML modeling .
Approach: They propose a multimodal instruction data evolution framework that iteratively enhances data quality through fine-grained perception, cognitive reasoning, and interaction evolution.
Outcome: The proposed approach improves MLLM performance in nine vision-language tasks while using significantly less data.
From What to Why: Improving Relation Extraction with Rationale Graph (2021.findings-acl)

Copied to clipboard

Challenge: Existing neural relation extraction models are limited by entity type and textual context.
Approach: They propose a novel RAtionale Graph to organize co-occurrence constraints among entity types, triggers and relations in a holistic graph view.
Outcome: The proposed method outperforms baselines significantly and achieves state-of-the-art performance on document-level and sentence-level RE benchmarks.
DUAL RM: Beyond Rule-based Preference Reward Modeling via Meta-Reward (2026.acl-long)

Copied to clipboard

Challenge: Existing preference-based reward modeling methods face a recursive dependency where each verifier requires a meta-verifier, leading to continuous and costly dependence on human annotation.
Approach: They propose a dual RM that couples discriminative and generative reward models under a non-parametric meta-reward.
Outcome: The proposed model achieves strong performance across major preference benchmarks and even when trained exclusively on language modality, it exhibits robust cross-modal transfer on Omni-RewardBench.
Unleashing LLM Reasoning Capability via Scalable Question Synthesis from Scratch (2025.acl-long)

Copied to clipboard

Challenge: Existing methods to improve the mathematical reasoning capabilities of Large Language Models (LLMs) are limited due to the proprietary nature of the data.
Approach: They propose a data synthesis method that generates large-scale mathematical reasoning datasets using lightweight 7B-scale models.
Outcome: The proposed method outperforms existing open-source datasets in both in-domain and out-of-domain evaluations and shows improvements in code reasoning tasks.
ReflectEvo: Improving Meta Introspection of Small LLMs by Learning Self-Reflection (2025.findings-acl)

Copied to clipboard

Challenge: ReflectEvo-460k is a large-scale, comprehensive, self-generated reflection dataset with broadened instructions and diverse multi-domain tasks.
Approach: They propose a pipeline that iteratively generates self-reflection for self-training and a large-scale reflection dataset with broadened instructions and diverse multi-domain tasks.
Outcome: The proposed pipeline improves Llama-3 reasoning ability by up to 71.2% and Mistral by upto 44.4%.
Neural Relation Classification with Text Descriptions (C18-1)

Copied to clipboard

Challenge: State-of-the-art methods for relation classification suffer from data sparsity issue greatly.
Approach: They propose a new neural relation classification method which integrates entities’ text descriptions into deep neural networks models.
Outcome: The proposed method achieves much better experimental results than other state-of-the-art methods on the SemEval 2010 dataset.
Generative Reward Modeling via Synthetic Criteria Preference Learning (2025.acl-long)

Copied to clipboard

Challenge: Generative Reward Models (GenRMs) leverage synthesized Chains of Thought (CoT) but this approach introduces risks of overoptimization due to the inability to guarantee the correctness of the CoTs.
Approach: They propose a criteria-based preference tree for GenRMs that uses chain of thought to generate reasoning . they show that synthesized data can be learned using a long CoT format .
Outcome: The proposed model shows significant improvements over baselines on multiple human preference benchmarks.
One-Shot Learning as Instruction Data Prospector for Large Language Models (2024.acl-long)

Copied to clipboard

Challenge: Contemporary practices in instruction tuning often hinge on enlarging data scaling without a clear strategy for ensuring data quality.
Approach: They propose a method that leverages one-shot learning to discern and select high-quality instruction data from extensive datasets.
Outcome: Nuggets outperforms existing methods on MT-Bench and Alpaca-Eval benchmarks.
A Contrastive Framework for Learning Sentence Representations from Pairwise and Triple-wise Perspective in Angular Space (2022.acl-long)

Copied to clipboard

Challenge: Existing methods for learning sentence representations focus on constitution of positive and negative representation pairs and do not focus on training objective.
Approach: They propose a new method to learn sentence representations using BERT-like pre-trained models . they use a pairwise discriminating power and a model to model the entailment relation of triplet sentences .
Outcome: The proposed method outperforms the previous state-of-the-art on diverse sentence related tasks.
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.
OFFSIDE: Benchmarking Unlearning Misinformation in Multimodal Large Language Models (2026.findings-acl)

Copied to clipboard

Challenge: Existing benchmarks for MU are limited by a lack of image diversity and coarse-grained unlearning targets.
Approach: They propose a benchmark to evaluate misinformation unlearning in MLLMs . OFFSIDE supports advanced unlearning targets such as fine-grained unlearning and visual rumor removal.
Outcome: OFFSIDE supports advanced unlearning targets, such as fine-grained unlearning and visual rumor removal.
Document-level Relation Extraction with Dual-tier Heterogeneous Graph (2020.coling-main)

Copied to clipboard

Challenge: Existing methods focus on extracting relations from single sentence . document-level relation extraction requires a comprehension of the whole document .
Approach: They propose a graph-based model with Dual-tier Heterogeneous Graph (DHG) for document-level relation extraction.
Outcome: The proposed model achieves state-of-the-art performance on two widely used datasets.
Tool learning via Inference-time Scaling and Cycle Verifier (2025.findings-acl)

Copied to clipboard

Challenge: In inference-time scaling, Chain-of-Thought (CoT) data is scarce or even unavailable.
Approach: They propose a method which establishes an inference cycle to synthesize user queries and CoT data.
Outcome: The proposed method achieves a 75.4% pass rate and a 79.6% win rate using small models in StableToolBench.
Unlocking Recursive Thinking of LLMs: Alignment via Refinement (2025.findings-acl)

Copied to clipboard

Challenge: Existing methods for recursive reasoning are limited due to lack of expert-curated data.
Approach: They propose a method that unlocks the potential of Large Language Models for recursive reasoning through long-form Chain of Thought.
Outcome: The proposed method outperforms preference optimization methods on the openAI o1-series models by 20% on 3k synthetic samples.
Escaping the Echo Trap: On Credit Assignment Failure in Multi-turn LLM Self-Reflection (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for multi-turn self-reflection are limited by the Echo Trap problem . the model is limited by its inherent capabilities and repeats earlier reflections to preserve reward signals .
Approach: They propose a tree-structured extension of GRPO for multi-turn self-reflection which enables more accurate advantage estimation.
Outcome: The proposed method mitigates behavior collapse and improves performance across benchmarks.
From Awareness to Adaptability: Enhancing Tool Utilization for Scientific Reasoning (2025.findings-acl)

Copied to clipboard

Challenge: Existing approaches enhance reasoning through Chain-of-Thought, Program-ofThough, and Tool-Integration.
Approach: They propose a tool-awareness training method that leverages both forward and backward data generation strategies to strengthen the model’s conscious and selective tool utilization in multi-step reasoning tasks.
Outcome: The proposed method improves the model's tool utilization capabilities, including proactivity and execution success rates.
Open-ended Long Text Generation via Masked Language Modeling (2023.acl-long)

Copied to clipboard

Challenge: Pre-trained autoregressive language models have dominated OPen-ended Long Text Generation (Open-LTG) however, the low inference efficiency of AR impedes their usability.
Approach: They propose a representative iterative non-autoregressive (NAR) decoding strategy to improve inference efficiency for Open-LTG.
Outcome: The proposed model can generate short text and collapse for long text modeling.

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