Papers by Pengfei Zhou

26 papers
GroupToM-Bench: Benchmarking Group Theory of Mind and Nonlinear Social Emergence in MLLMs (2026.acl-long)

Copied to clipboard

Challenge: Existing models for general intelligence fail to model how mental states interact and crystallize into group-level outcomes.
Approach: They propose a multimodal benchmark for group-level Theory of Mind (ToM) to probe nonlinear collective behavior.
Outcome: The proposed model performs significantly below human levels, exposing blind spots in modeling social structures and nonlinear collective behavior.
MARS2: Scaling Multi-Agent Tree Search via Reinforcement Learning for Code Generation (2026.acl-long)

Copied to clipboard

Challenge: Existing approaches to reinforcement learning are decoupled from structured search due to limited trajectory diversity.
Approach: They propose a unified RL framework that integrates multiple agents within a shared tree-structured search environment.
Outcome: Experiments show that MARS2 improves performance across diverse model combinations and training settings.
CogKTR: A Knowledge-Enhanced Text Representation Toolkit for Natural Language Understanding (2022.emnlp-demos)

Copied to clipboard

Challenge: Existing knowledge-enhanced methods are limited to knowledge-intensive tasks.
Approach: They propose a knowledge-enhanced text representation toolkit for natural language understanding . it combines knowledge acquisition, knowledge representation, knowledge injection and knowledge application .
Outcome: The proposed toolkit supports knowledge acquisition, knowledge representation, knowledge injection, and knowledge application.
LLM-Based Human-Agent Collaboration and Interaction Systems: A Survey (2026.findings-acl)

Copied to clipboard

Challenge: Recent advances in large language models (LLMs) have sparked growing interest in building fully autonomous agents.
Approach: They propose to integrate human-provided information, feedback, or control into the agent system to enhance system performance, reliability, and safety.
Outcome: The proposed systems improve system performance, reliability, and safety by integrating human-provided information, feedback, or control into the agent system.
Dissecting Human and LLM Preferences (2024.acl-long)

Copied to clipboard

Challenge: a recent study shows that human and Large Language Model preferences are important for model fine-tuning and evaluation.
Approach: They dissect the preferences of human and 32 different Large Language Models to understand their quantitative composition.
Outcome: The proposed model is compared with 32 different large language models using real-world user-model conversations.
LongCLI-Bench: A Preliminary Benchmark and Study for Long-horizon Agentic Programming in Command-Line Interfaces (2026.findings-acl)

Copied to clipboard

Challenge: Existing benchmarks for agentic programming in long-horizon command-line interface tasks are limited by short task horizons, data contamination from GitHub scraping, and a lack of fine-grained evaluation metrics.
Approach: They propose a benchmark to evaluate agentic capabilities across long-horizon command-line interface tasks.
Outcome: The proposed benchmarks cover four engineering categories: from scratch, feature addition, bug fixing, and refactoring.
Multi-Dimensional Evaluation of Text Summarization with In-Context Learning (2023.findings-acl)

Copied to clipboard

Challenge: In-context learning-based evaluators are competitive with learned evaluation frameworks for text summarization tasks.
Approach: They propose to use large language models as multi-dimensional evaluators using in-context learning to evaluate text summarization tasks.
Outcome: The proposed frameworks are competitive with existing frameworks on relevance and factual consistency, the authors show .
Diversity-oriented Data Augmentation with Large Language Models (2025.acl-long)

Copied to clipboard

Challenge: Existing data augmentation methods focus on increasing sample numbers while neglecting sample distribution diversity, which can lead to model overfitting.
Approach: They propose a data augmentation framework that focuses on sample distribution diversity and trains a large language model as a diverse paraphraser.
Outcome: The proposed framework achieves an average performance gain of 10.52% surpassing the runner-up baseline with more than three percentage points.
AscendKernelGen: LLM-Driven Kernel Generation for NPUs (2026.findings-acl)

Copied to clipboard

Challenge: Neural Processing Units (NPUs) are critical for AI infrastructure, but their development remains a bottleneck due to vendor-specific Domain-Specific Languages (DSLs).
Approach: They propose a framework for NPU kernel development that bridges the gap in hardware-specific coding . compiler success on complex Level-2 kernels improves from 0% to 95.5%, they say .
Outcome: The proposed framework bridges the gap in hardware-specific coding, showing a near-zero success rate on complex kernels.
MPBench: A Comprehensive Multimodal Reasoning Benchmark for Process Errors Identification (2025.findings-acl)

Copied to clipboard

Challenge: Existing benchmarks of large language models focus on error detection, neglecting other scenarios like reasoning search.
Approach: et al. propose a multi-task, multimodal benchmark to assess effectiveness of PRMs . step correctness, answers aggregation and reasoning process search are evaluated . ethical principles of MPBench are based on a set of evaluation paradigms based in a text-based benchmark .
Outcome: a new benchmark assesses the effectiveness of large language models (LLMs) in multiple scenarios . it uses three evaluation paradigms to assess the effectiveness and compares them with existing models . a the proposed model improves reasoning accuracy by providing stepwise feedback for multi-step reasoning results .
Retrieval Heads are Dynamic (2026.acl-long)

Copied to clipboard

Challenge: Recent studies have identified "retrieval heads" in Large Language Models responsible for extracting information from input contexts.
Approach: They propose to examine retrieval heads from a dynamic perspective . they establish that retrieval head activation is highly dynamic and functionally irreplaceable .
Outcome: The proposed model's hidden state encodes a predictive signal for future retrieval head patterns, indicating an internal planning mechanism.
Ambiguity Awareness Optimization: Towards Semantic Disambiguation for Direct Preference Optimization (2025.emnlp-main)

Copied to clipboard

Challenge: Direct Preference Optimization (DPO) is a widely used reinforcement learning from human feedback (RLHF) method across various domains.
Approach: They propose an approach that automatically re-weights ambiguous content to reduce ambiguities by calculating semantic similarity from preference pairs.
Outcome: The proposed approach outperforms state-of-the-art approaches in performance across multiple model scales and widely adopted benchmark datasets.
Automatic ICD Coding via Interactive Shared Representation Networks with Self-distillation Mechanism (2021.acl-long)

Copied to clipboard

Challenge: Existing methods for ICD coding ignore the long-tail of code frequency or noisy clinical notes.
Approach: They propose to use an interactive shared representation network to model code co-occurrences while focusing on the clinical note's noteworthy part and extract valuable information through a self-distillation learning mechanism to solve the long-tail problem.
Outcome: The proposed model reduces the long-tail of code frequency and noise in clinical notes and extracts valuable information through a self-distillation learning mechanism.
DTELS: Towards Dynamic Granularity of Timeline Summarization (2025.naacl-long)

Copied to clipboard

Challenge: Existing timeline summarizations lack flexibility to meet diverse granularity needs . a fine-grained timeline showing the technical details is preferred for news topics .
Approach: They propose a new paradigm to construct adaptive timelines based on user instructions or requirements.
Outcome: The proposed timelines are informative and granularly consistent, but they struggle to generate consistent timelines.
From Signal Degradation to Computation Collapse: Uncovering the Two Failure Modes of LLM Quantization (2026.findings-acl)

Copied to clipboard

Challenge: Existing research on PTQ spans three primary directions.
Approach: They conduct a systematic analysis of post-training quantization failures using PTQ . they show that targeted repair can mitigate Signal Degradation but remains ineffective for Computation Collapse .
Outcome: The proposed method mitigates Signal Degradation but remains ineffective for Computation Collapse.
When KV Cache Reuse Fails in Multi-Agent Systems: Cross-Candidate Interaction is Crucial for LLM Judges (2026.acl-long)

Copied to clipboard

Challenge: Multi-agent LLMs generate multiple candidate responses that are aggregated by an LLM judge.
Approach: They propose to advocate KV cache reuse across partially shared contexts and report substantial speedups for generation agents.
Outcome: The proposed reuse strategies weaken cross-candidate attention, especially for later candidate blocks, and highlight judge-centric inference as a distinct regime that requires dedicated, risk-aware system design.
Task-Stratified Knowledge Scaling Laws for Post-Training Quantized Large Language Models (2026.findings-acl)

Copied to clipboard

Challenge: Existing scaling laws focus on general performance, overlooking crucial fine-grained factors and how quantization differentially impacts diverse knowledge capabilities.
Approach: They propose a framework that unifies model size, bit-width, and fine-grained factors into memorization, application, and reasoning.
Outcome: The proposed framework shows strong fit and cross-architecture consistency on 293 different PTQ configurations.
Lightweight Haar Wavelet Subband Pruning for LLMs (2026.findings-acl)

Copied to clipboard

Challenge: Large language models (LLMs) have impressive performance but require computational and memory resources.
Approach: They propose a post-training framework that uses a Haar wavelet transform to prune weights.
Outcome: The proposed pruning framework reduces pruning time and computational costs by removing less important weights while preserving model architecture.
AIMMerging: Adaptive Iterative Model Merging Using Training Trajectories for Language Model Continual Learning (2025.emnlp-main)

Copied to clipboard

Challenge: Recent model merging-based methods struggle to effectively manage the trade-off between learning new knowledge and preventing catastrophic forgetting.
Approach: They propose a model merging framework that utilizes learning and forgetting signals from the training trajectory to dynamically monitor the model’s training status.
Outcome: The proposed framework achieves significant performance improvements over existing state-of-the-art methods on three CL benchmarks with various model sizes (from 770M to 13B).
M2Edit: Locate and Edit Multi-Granularity Knowledge in Multimodal Large Language Model (2025.emnlp-main)

Copied to clipboard

Challenge: Existing knowledge editing methods for MLLMs lack multi-granularity knowledge . existing knowledge editing approaches lack multimodality knowledge and generalize to multimodal data.
Approach: They propose a multimodal knowledge editing method which integrates key knowledge layers within MLLMs and collaboratively edits them.
Outcome: The proposed method improves visual generality performance on knowledge data of different granularities.
LASS: A Novel and Economical Data Augmentation Framework Based on Language Models for Debiasing Opinion Summarization (2025.coling-main)

Copied to clipboard

Challenge: Existing methods to generate negative summaries are expensive and lack the capacity to generate large data sets.
Approach: They propose a data augmentation framework based on LArge and Small language models for debiaSing opinion summarization that generates a small number of synthesized negative reviews by rewriting the positive text via a large language model.
Outcome: The proposed framework can generate large numbers of negative reviews by rewriting the positive text using a large language model and training a disentangle reconstruction model based on the generated data.
SepSeq: A Training-Free Framework for Long Numerical Sequence Processing in LLMs (2026.findings-acl)

Copied to clipboard

Challenge: Existing large-scale large-context models suffer from performance degradation when processing long numerical sequences.
Approach: They propose a framework to mitigate attention dispersion by strategically inserting separator tokens into the model to recalibrat attention to local segments while preserving global context.
Outcome: The proposed framework improves accuracy and reduces inference token consumption by 16.4% on 9 widely-adopted LLMs.
DavIR: Data Selection via Implicit Reward for Large Language Models (2025.acl-long)

Copied to clipboard

Challenge: 6% of Alpaca dataset selected with DavIR can steer both LLaMA and Gemma models to produce superior performance compared to the same models trained on the full 52K dataset.
Approach: They propose a model-based data selection method for post-training Large Language Models . they generalize Reducible Holdout Loss to core-set selection problem of causal language modeling .
Outcome: The proposed method can steer both LLaMA and Gemma models to superior performance compared to the same models trained on the full 52K dataset.
Opinions Are Not Always Positive: Debiasing Opinion Summarization with Model-Specific and Model-Agnostic Methods (2024.lrec-main)

Copied to clipboard

Challenge: Existing opinion summarization frameworks are reluctant to generate negative summaries given input of negative opinions.
Approach: They propose to disentangle input into sentiment-relevant and sentiment-irrelevant components through adversarial loss.
Outcome: The proposed approaches reduce sentiment bias in the existing opinion summarization dataset . the proposed approaches generate better summaries with a more balanced emotional polarity distribution .
Benchmarking Temporal Reasoning and Alignment Across Chinese Dynasties (2026.eacl-short)

Copied to clipboard

Challenge: Existing temporal reasoning benchmarks rely on rule-based construction and lack contextual depth . a recent study found existing LLMs struggle with nuanced temporal understanding .
Approach: a benchmark is designed to evaluate LLMs on temporal reasoning in Chinese dynasties.
Outcome: a new benchmark evaluates LLMs on temporal reasoning across Chinese dynasties . it emphasizes cross-entity relationships, pairwise temporal alignment, contextualized and culturally-grounded reasoning . results show existing LLM benchmarks struggle with nuanced temporal understanding .

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