Papers by Yi Pan

11 papers
AutoMixAlign: Adaptive Data Mixing for Multi-Task Preference Optimization in LLMs (2025.acl-long)

Copied to clipboard

Challenge: Existing approaches to align large language models rely on large ablation studies, heuristics, or human intuition to produce models with strong performance across tasks.
Approach: They propose an algorithm that mixes datasets during LLM training to balance performance across multiple tasks.
Outcome: The proposed algorithm outperforms existing methods on multitask alignment setups and achieves convergence rate of O(1/T) in the convex case.
HELENE: Hessian Layer-wise Clipping and Gradient Annealing for Accelerating Fine-tuning LLM with Zeroth-order Optimization (2025.emnlp-main)

Copied to clipboard

Challenge: Large language models (LLMs) face memory challenges due to the high cost of backpropagation.
Approach: They propose a zeroth-order (ZO) optimization that matches memory usage to inference . they propose scalable and memory-efficient zeroth order (ZE) optimizer that integrates annealed A-GNB gradients with diagonal Hessian estimation and layer-wise clipping as a second-order pre-conditioner.
Outcome: The proposed algorithm outperforms state-of-the-art methods with an average speedup of 20 over MeZO on RoBERTa-large and OPT-1.3B.
Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failures (2026.acl-long)

Copied to clipboard

Challenge: Recent research has shown that reinforcement learning can elicit intriguing emergent reasoning behaviors.
Approach: They propose a comprehensive survey of the mechanistic understanding of large reasoning models . they organize findings into three core dimensions: 1) training dynamics, 2) reasoning mechanisms, and 3) unintended behaviors.
Outcome: This paper synthesizes the mechanistic understanding of large reasoning models into three dimensions . authors outline a roadmap for future studies including improved interpretability and methodologies .
Alexa Conversations: An Extensible Data-driven Approach for Building Task-oriented Dialogue Systems (2021.naacl-demos)

Copied to clipboard

Challenge: Traditional goal-oriented dialogue systems require annotations which are hard to obtain for every new domain, limiting scalability.
Approach: They propose a data-driven approach to building goal-oriented dialogue systems . they use a seed dialogue simulator to generate annotated conversations instead of collecting annotations .
Outcome: The proposed system improves turn-level action signature prediction accuracy by 50% . the system is scalable, extensible and data efficient .
ProPy: Building Interactive Prompt Pyramids upon CLIP for Partially Relevant Video Retrieval (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing models for PRVR use unimodal features, but powerful pretrained vision-language models like CLIP are underexplored.
Approach: ProPy is a model with systematic architectural adaptation of CLIP specifically designed for PRVR.
Outcome: ProPy outperforms existing models on three public datasets in terms of performance on the datasets.
RESIN: A Dockerized Schema-Guided Cross-document Cross-lingual Cross-media Information Extraction and Event Tracking System (2021.naacl-demos)

Copied to clipboard

Challenge: We present a new information extraction system that can construct temporal event graphs from news documents.
Approach: They propose a temporal event graph extraction system that can extract news documents . they extend the system from sentence-level event extraction to cross-document cross-media event extraction .
Outcome: The proposed system can extract temporal event graphs from news documents in multiple languages and multiple data modalities.
Feedback to Reasoning: LLM-Assisted Molecular Optimization with Domain Feedback and Historical Reasoning (2026.findings-acl)

Copied to clipboard

Challenge: Existing methods for molecular optimization do not leverage domain feedback and historical knowledge with reasoning traces and chemical insights.
Approach: They propose a conversational molecular optimization pipeline that enables LLMs to accumulate and retrieve past actions, rationales, and feedback.
Outcome: The proposed framework transforms LLMs from passive text generators into agentic experts that learn both actions and reasoning from experience.
CoTKR: Chain-of-Thought Enhanced Knowledge Rewriting for Complex Knowledge Graph Question Answering (2024.emnlp-main)

Copied to clipboard

Challenge: Existing knowledge rewriting methods may include irrelevant information, omit crucial details, or fail to align with the question’s semantics.
Approach: They propose a new rewriting method CoTKR for generating reasoning traces and corresponding knowledge in an interleaved manner, thereby mitigating the limitations of single-step knowledge rewrite.
Outcome: The proposed method mitigates the limitations of single-step knowledge rewriting and bridges the preference gap between the knowledge reactor and the question answering (QA) model.
VillagerAgent: A Graph-Based Multi-Agent Framework for Coordinating Complex Task Dependencies in Minecraft (2024.findings-acl)

Copied to clipboard

Challenge: Multi-agent collaboration using LLMs is a challenging research topic that aims to enable multiple autonomous agents to coordinate their actions and achieve a common goal.
Approach: They propose a benchmark for multi-agent collaboration in the Minecraft environment and introduce a Directed Acyclic Graph Multi-Agent Framework to resolve complex inter-ag dependencies.
Outcome: The proposed framework outperforms existing ModelVerse, reducing hallucinations and improving task decomposition efficacy.
FlattenQuant: Breaking through the Inference Compute-bound for Large Language Models with Per-tensor Quantization (2024.lrec-main)

Copied to clipboard

Challenge: Large language models (LLMs) have demonstrated state-of-the-art accuracies across tasks, but their latency and GPU memory consumption limit their performance.
Approach: They propose a method which flattens the tensor to achieve low bit per-tensori quantization with minimal accuracy loss.
Outcome: The proposed method achieves low bit per-tensor quantization with minimal accuracy loss.
Synapse: Empowering LLM Agents with Episodic-Semantic Memory via Spreading Activation (2026.findings-acl)

Copied to clipboard

Challenge: Large Language Models excel at generalized reasoning, but lack the ability to accumulate experiences and maintain narrative coherence over long horizons.
Approach: They propose a unified memory architecture that transcends static vector similarity.
Outcome: The proposed model outperforms state-of-the-art methods in temporal and multihop reasoning tasks.

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