China

This paper identifies that misinterpreting the context can be a significant issue during the reasoning process of large language models, spanning from smaller models like Llama3.2-3B-Instruct to cutting-edge ones like DeepSeek-R1. We introduce a novel, post-training approach called **Stick to the Facts (SIFT)** to tackle this. SIFT leverages increasing inference-time compute to ground LLM reasoning in contexts. At the core of SIFT lies the Sticker, which is generated by the model itself to explicitly emphasize the key information within the context. Given the Sticker, SIFT generates two predictions—one from the Sticker alone and one from the query augmented with the Sticker. If they differ, the Sticker is sequentially refined via forward optimization (to better align the extracted facts with the query) and inverse generation (to conform with the model’s inherent tendencies) for more faithful reasoning outcomes. Studies across diverse models (from 3B to 100B+) and benchmarks (e.g., MATH, AIME) reveal consistent performance improvements. Notably, SIFT improves the pass@1 accuracy of DeepSeek-R1 on AIME2024 from 78.33% to **85.67%** and that on AIME2025 from 69.8% to **77.33%**. Code will be public after acceptance.

EMNLP 2025

SIFT: Grounding LLM Reasoning in Contexts via Stickers

llms

reasoning

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community."<br>

*Message from the General Chair, Dirk Hovy*





<br>

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

poster

Unlearning aims to remove copyrighted, sensitive, or private content from large language models (LLMs) without a full retraining. In this work, we develop a multi-task unlearning benchmark LUME that features three tasks: (1) unlearn synthetically generated creative short novels, (2) unlearn synthetic biographies with sensitive information, and (3) unlearn a collection of public biographies. We further release two fine-tuned LLMs of 1B and 7B parameter sizes as the target models. We conduct detailed evaluations of several recently-proposed algorithms and present results on carefully crafted metrics to understand their behavior and limitations.

LUME: LLM Unlearning with Multitask Evaluations

We investigate the effect of problem presentation on LLMs' ability to solve optimization problems. We introduce the dataset of Everyday Hard Optimization Problems (EHOP), a collection of NP-hard problems expressed in natural language. EHOP includes problem formulations that could be found in computer science textbooks (e.g., graph coloring), versions that are dressed up as problems that could arise in real life (e.g., party planning), and variants with inverted rules. We find that state-of-the-art LLMs, across multiple prompting strategies, systematically solve textbook problems more accurately than their real-life and inverted counterparts. While reasoning models are more robust, they nonetheless show high variance across problem presentations, suggesting they lack a truly robust reasoning mechanism. We argue that this constitutes evidence that LLMs are still heavily dependent on what was seen in training, and struggle to generalize to novel problems.

A Knapsack by Any Other Name: Presentation impacts LLM performance on NP-hard problems

Persona agents, which are LLM agents conditioned to act according to an assigned persona, enable contextually rich and user aligned interactions across domains like education and healthcare. However, evaluating how faithfully these agents adhere to their personas remains a significant challenge, particularly in free-form settings that demand consistency across diverse, persona-relevant environments. We introduce PersonaGym, the first dynamic evaluation framework for persona agents, and PersonaScore, a human-aligned automatic metric grounded in decision theory that enables comprehensive, large-scale evaluation. Our evaluation of 10 leading LLMs across 200 personas and 10,000 questions reveals significant advancement opportunities. For example, GPT-4.1 had the exact same PersonaScore as LLaMA-3-8b despite being a more recent and advanced closed source model. Importantly, increased model size and complexity do not necessarily enhance persona agent capabilities, underscoring the need for algorithmic and architectural innovation toward faithful, performant persona agents. Code available at: anonymous.4open.science/r/PersonaGym-014A

PersonaGym: Evaluating Persona Agents and LLMs

We propose Rotate, Clip, and Partition (RCP), a Quantization-Aware Training (QAT) approach that first realizes extreme compression of LLMs with W2A4KV4 (2-bit weight, 4-bit activation, and 4-bit KV-cache) configuration. RCP integrates recent rotation techniques with a novel non-uniform weight quantizer design by theoretically and empirically analyzing the impact of rotation on the non-uniformity of weight distribution. Our weight quantizer, Learnable Direct Partitioning (LDP), introduces learnable parameters to directly learn non-uniform intervals jointly with LLM weights. We also present a GPU kernel supporting GEMV on non-uniform W2A4 as proof of concept. Experiments show that RCP can compress LLaMA-2-7B to W2A4KV4 with a loss of only 2.84 WikiText2 PPL and 5.29 times reduced memory footprint. Furthermore, RCP can quantize challenging mobile-targeted LLaMA-3.2 models and domain-specific WizardCoder-7B and MetaMath-7B with no critical problems such as convergence failure and repetition. Code will be made available at \url{blind_review}.

Rotate, Clip, and Partition: Towards W2A4KV4 Quantization by Integrating Rotation and Learnable Non-uniform Quantizer

Retrieval-augmented generation (RAG) enables large language models (LLMs) to address queries beyond their internal knowledge by integrating domain knowledge in specialized corpus, which necessitates the generation of benchmarks on specific corpus to evaluate RAG systems. However, existing automated generation methods exhibit Weak Applicability and Weak Scalability. Weak Applicability refers to the reliance on metadata from specific corpora for query generation, constraining applicability to other corpora. Weak Scalability is characterized by fixed query content after generation, unable to dynamically increase difficulty, limiting scalability of the query. To overcome these issues, we propose AutoEvolve, an applicable approach for dynamically evolving queries to construct scalable RAG benchmarks. Our approach is grounded in three key innovations: (i) a corpus-agnostic method for constructing the universal entity-document graph; (ii) a suite of evolution operations designed to dynamically update queries; and (iii) a difficulty-guided metric that directs query evolution process. Through experiments on three generated benchmarks, we demonstrate that AutoEvolve evolves queries that are significantly more challenging, paving the way for more applicable and scalable RAG evaluations.

AutoEvolve: Automatically Evolving Queries for Applicable and Scalable Retrieval-Augmented Generation Benchmarking

Knowledge editing aims to update LLMs with new information without costly retraining. However, consistently reflecting these updates in complex multi-hop QA, which demands reasoning over interconnected facts, is challenging. Many existing methods overlook the interplay with pre-existing knowledge, leading to inconsistent edit propagation. To overcome this, we introduce StepKE (Stepwise Knowledge Editing for Multi-hop QA), a novel framework for robustly integrating edited and existing knowledge for coherent multi-hop reasoning. StepKE uniquely decomposes multi-hop questions into sequential single-hop sub-questions, retrieving relevant facts (both edited and pre-existing) from an external knowledge graph for each step. Crucially, it employs context-aware prompting with prior reasoning history and an LTE-inspired fine-tuning for precise edit propagation. This systematic integration enables effective stepwise reasoning. Experiments show StepKE generates significantly more accurate and consistent responses than baselines, showcasing strong knowledge editing and integration in multi-hop QA.

StepKE: Stepwise Knowledge Editing for Multi-Hop Question Answering

We study Attributed Question Answering (abbr., AQA), a newly-released long-form answer generation task. The tailored and efficient training programmes haven't yet been leveraged to strengthen AQA models. This hinders the simultaneous enhancement of their essential capabilities, including evidence identification, cross-source relation recognition and anti-distraction reasoning. To address the issue, we propose a tailored progressive curriculum learning approach, and use it to optimize both encoder-decoder and decoder-only AQA models. Experiments on the benchmark QuoteSum show that our approach yields substantial improvements and enables the AQA performance to reach 73.9% Sem-F1 score.

Enhancing Attributed Question Answering using Tailored Progressive Curriculum Learning

We introduce COMI-LINGUA, the largest manually annotated Hindi-English code-mixed dataset, comprising 125K+ high-quality instances across five core NLP tasks: Matrix Language Identification, Token-level Language Identification, POS Tagging, Named Entity Recognition, and Machine Translation. Each instance is annotated by three bilingual annotators, yielding over 376K expert annotations with strong inter-annotator agreement (Fleiss' Kappa >= 0.81). The rigorously preprocessed and filtered dataset covers both Devanagari and Roman scripts, and spans diverse domains such as social media text, news, and formal communications, ensuring real-world linguistic coverage. Evaluation reveals that closed-source LLMs significantly outperform traditional tools and open-source models. Notably, one-shot prompting consistently boosts performance across tasks, especially in structure-sensitive predictions like POS and NER, highlighting the effectiveness of prompt-based adaptation in code-mixed, low-resource settings. COMI-LINGUA is publicly available at: \url{https://anonymous.4open.science/r/CodeMixing/}.

COMI-LINGUA: Expert Annotated Large-Scale Dataset for Multitask NLP in Hindi-English Code-Mixing

Retrieval-Augmented Generation (RAG) has become essential for large-scale code generation, grounding predictions in external code corpora to improve factuality. However, a critical yet underexplored aspect of RAG pipelines is chunking---the process of dividing documents into retrievable units. Existing line-based chunking heuristics often break semantic structures, splitting functions or merging unrelated code, which can degrade generation quality. We propose chunking via Abstract Syntax Trees (cAST), a structure-aware method that recursively breaks large AST nodes into smaller chunks and merges sibling nodes while respecting size limits. This approach generates self-contained, semantically coherent units across programming languages and tasks, improving performance on diverse code generation tasks, e.g., boosting Recall@5 by 4.3 points on RepoEval retrieval and Pass@1 by 2.67 points on SWE-bench generation. Our work highlights the importance of structure-aware chunking for scaling retrieval-enhanced code intelligence.

cAST: Enhancing Code Retrieval-Augmented Generation with Structural Chunking via Abstract Syntax Tree

Our goal is to build language model based multi-agent systems for complex planning problems involving multiple explicit and implicit constraints, some of which may be commonsense. Our initial investigations reveal that large language models (LLMs) are often unable to maintain consistency across the planning process, whereas large reasoning models (LRMs) struggle with handling implicit commonsense constraints. In response, we introduce LRPlan, a novel domain-independent, language-based multi-agent architecture where LLM and LRM-based agents collaborate at training time to abstract important patterns, heuristics and insights about the domain. At test time, they collaborate in implementing these learned patterns and insights for a new planning instance. We perform experiments on two domains, TravelPlanner and TimeArena-Static, and use two LLM-LRM combinations from GPT and DeepSeek families. We find that LRPlan outperforms various multi-agent and single-agent baselines obtaining notably higher accuracy as well as cost efficiency.

Downloads

Next from EMNLP 2025

LUME: LLM Unlearning with Multitask Evaluations

Similar lecture

Sticker-TTS: Learn to Utilize Historical Experience with a Sticker-driven Test-Time Scaling Framework