Papers by Yifeng Wang

12 papers
Stop Looking for “Important Tokens” in Multimodal Language Models: Duplication Matters More (2025.emnlp-main)

Copied to clipboard

Challenge: Vision tokens in multimodal large language models often dominate computational overhead due to excessive length compared to linguistic modality.
Approach: They propose a token pruning method which defines an importance criterion for vision tokens and prunes the unimportant vision token during inference.
Outcome: The proposed method can prune 88.9% of vision tokens while maintaining comparable performance.
Stability Implies Redundancy: Delta Attention Selective Halting for Efficient Long-Context Prefilling (2026.acl-long)

Copied to clipboard

Challenge: Existing methods to reduce sequence length rely on heuristics that break compatibility with hardware-efficient kernels like FlashAttention.
Approach: They propose a method that selectively halts stabilized tokens by monitoring layer-wise update dynamics of the self-attention mechanism.
Outcome: The proposed method can reduce prefill complexity while preserving model accuracy and hardware efficiency.
FuzzAug: Data Augmentation by Coverage-guided Fuzzing for Neural Test Generation (2025.findings-emnlp)

Copied to clipboard

Challenge: Using large language models to generate meaningful tests is expensive and time-consuming .
Approach: They propose a data augmentation technique that incorporates valid testing semantics and diverse coverage-guided inputs into large language models.
Outcome: The proposed technique improves performance over the baselines by incorporating valid testing semantics and providing diverse coverage-guided inputs.
Beyond Logits: Aligning Feature Dynamics for Effective Knowledge Distillation (2025.acl-long)

Copied to clipboard

Challenge: Knowledge distillation (KD) compresses large language models into lightweight versions called student models.
Approach: They propose to align the entire feature dynamics between teacher and student models by using two additional loss terms to achieve this.
Outcome: The proposed method matches the entire feature dynamics between teacher and student models rather than just the final states.
NOVA: An Iterative Planning Framework for Enhancing Scientific Innovation with Large Language Models (2025.findings-acl)

Copied to clipboard

Challenge: Existing approaches to generate research ideas rely on retrieval or prompt engineering to generate ideas.
Approach: They propose a method that uses iterative planning and search to boost creative potential of LLMs by integrating external knowledge with broader and deeper insights.
Outcome: The proposed method outperforms the current state-of-the-art in generating 2.5 times more top-rated ideas based on 170 seed papers in a Swiss Tournament evaluation.
Wukong-Reader: Multi-modal Pre-training for Fine-grained Visual Document Understanding (2023.acl-long)

Copied to clipboard

Challenge: Existing solutions for visual document understanding lack granularity of document textlines.
Approach: They propose a supervised pre-training program to leverage structural knowledge nested in document textlines to achieve fine-grained alignment between visual regions and texts.
Outcome: The proposed system performs better on various VDU tasks in English and Chinese.
MMErroR: A Benchmark for Erroneous Reasoning in Vision-Language Models (2026.acl-long)

Copied to clipboard

Challenge: Recent advances in vision-language models have improved performance in multi-modal learning.
Approach: They propose a multi-modal benchmark that embeds a single coherent reasoning error in 1997 samples.
Outcome: The proposed benchmark is based on a set of 1997 samples embedding a single coherent reasoning error.
Empowering Multi-Turn Tool-Integrated Agentic Reasoning with Group Turn Policy Optimization (2026.acl-long)

Copied to clipboard

Challenge: Current reinforcement learning methods suffer from coarse-grained, trajectory-level rewards that provide insufficient learning signals for complex multi-turn interactions, leading to training stagnation.
Approach: They propose a novel RL algorithm for training large language models for multi-turn tool-integrated reasoning (TIR) that incorporates three innovations: turn-level reward assignment that provides fine-grained feedback for individual turns, return-based advantage estimation where normalized discounted returns are calculated as advantages, and self-supervised reward shaping that exploits self-supervision signals from generated code to densify sparse binary outcome-based rewards.
Outcome: The proposed algorithm outperforms GRPO by 3.0% across diverse math reasoning benchmarks and improves grepo by 3.9% on commonsense reasoning and program synthesis tasks.
Chinese Inertial GAN for Handwriting Signal Generation and Recognition (2025.acl-long)

Copied to clipboard

Challenge: Inertial sensors can measure the acceleration and angular velocity of moving objects and are widely used in electronic devices such as smartphones, smartwatches, and fitness bands.
Approach: They propose to use Chinese glyph encoding, forced optimal transport, and semantic relevance alignment to acquire unlimited training samples for Chinese inertial writing recognition.
Outcome: The proposed system improves the performance of six widely used classifiers from 6.7% to 98.4%.
SyncThink: A Training-Free Strategy to Align Inference Termination with Reasoning Saturation (2026.findings-acl)

Copied to clipboard

Challenge: Large language models (LLMs) achieve strong reasoning with Chain-of-Thought prompting, but long and redundant traces substantially increase inference cost.
Approach: They propose a training-free and plug-and-play decoding method that reduces CoT overhead without modifying model weights.
Outcome: Experiments on GSM8K, MMLU, GPQA, and BBH show that SyncThink achieves 62.00% average Top@1 accuracy using 656 generated tokens and 28.68s latency, compared to 61.22%, 2141 tokens, and 92.01s for full CoT decoding.
Planning-Aware Code Infilling via Horizon-Length Prediction (2025.emnlp-main)

Copied to clipboard

Challenge: Current approaches to fill-in-the-middle (FIM) often fail to generate content that aligns well with the surrounding context.
Approach: They propose a training objective that teaches models to predict the number of remaining middle tokens at each step.
Outcome: The proposed training objective improves FIM performance by up to 24% on diverse benchmarks across file-level and repository-level.
Fusion or Defusion? Flexible Vision-and-Language Pre-Training (2023.findings-acl)

Copied to clipboard

Challenge: Existing approaches to vision-and-language pretraining (VLP) lack effectiveness and efficiency in downstream multimodal tasks.
Approach: They propose a flexible vision-and-language pre-training model by incorporating cross-modal fusions into a dual-encoder architecture and a cross-module knowledge transfer strategy to guide the training process.
Outcome: The proposed model is well-equipped with effectiveness and efficiency compared with other strong VLP models.

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