China

Faithful free-text explanations are important to ensure transparency in high-stakes AI decision-making contexts, but they are challenging to generate by language models and assess by humans. In this paper, we present a measure for Prediction-EXplanation (PEX) consistency, by extending the concept of weight of evidence. This measure quantifies how much a free-text explanation supports or opposes a prediction, serving as an important aspect of explanation faithfulness. Our analysis reveals that more than 62% explanations generated by large language models lack this consistency. We show that applying direct preference optimization improves the consistency of generated explanations across three model families, with improvement ranging from 43.1% to 292.3%. Furthermore, we demonstrate that optimizing this consistency measure can improve explanation faithfulness by up to 9.7%.

EMNLP 2025

A Necessary Step toward Faithfulness: Measuring and Improving Consistency in Free-Text Explanations

explanation faithfulness

natural language explanations

explainability

## 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

Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by retrieving relevant documents from external sources to improve factual accuracy and verifiability. However, this reliance introduces new attack surfaces within the retrieval pipeline, beyond the LLM itself. While prior RAG attacks have exposed such vulnerabilities, they largely rely on manipulating user queries, which is often infeasible in practice due to fixed or protected user inputs. This narrow focus overlooks a more realistic and stealthy vector: instructional prompts, which are widely reused, publicly shared, and rarely audited. Their implicit trust makes them a compelling target for adversaries to manipulate RAG behavior covertly. We introduce a novel attack for Adversarial Instructional Prompt (AIP) that exploits adversarial instructional prompts to manipulate RAG outputs by subtly altering retrieval behavior. By shifting the attack surface to the instructional prompts, AIP reveals how trusted yet seemingly benign interface components can be weaponized to degrade system integrity. The attack is crafted to achieve three goals: (1) naturalness, to evade user detection; (2) utility, to encourage use of prompts; and (3) robustness, to remain effective across diverse query variations. We propose a diverse query generation strategy that simulates realistic linguistic variation in user queries, enabling the discovery of prompts that generalize across paraphrases and rephrasings. Building on this, a genetic algorithm-based joint optimization is developed to evolve adversarial prompts by balancing attack success, clean-task utility, and stealthiness. Experimental results show that AIP achieves up to 95.23% attack success rate while preserving benign functionality. These findings uncover a critical and previously overlooked vulnerability in RAG systems, emphasizing the need to reassess the shared instructional prompts.

AIP: Subverting Retrieval-Augmented Generation via Adversarial Instructional Prompt

Ambiguity is pervasive in real-world questions, yet large language models (LLMs) often respond with confident answers rather than seeking clarification. In this work, we show that question ambiguity is linearly encoded in the internal representations of LLMs and can be both detected and controlled at the neuron level. During the model’s pre-filling stage, we identify that a small number of neurons, as few as one, encode question ambiguity information. Probes trained on these Ambiguity-Encoding Neurons (AENs) achieve strong performance on ambiguity detection and generalize across datasets, outperforming prompting-based and representation-based baselines. Layerwise analysis reveals that AENs emerge from shallow layers, suggesting early encoding of ambiguity signals in the model’s processing pipeline. Finally, we show that through manipulating AENs, we can control LLM's behavior from direct answering to abstention. Our findings reveal that LLMs form compact internal representations of question ambiguity, enabling interpretable and controllable behavior.

Sparse Neurons Carry Strong Signals of Question Ambiguity in LLMs

An embodied AI assistant operating on egocentric video must integrate spatial cues across time - for instance, determining where an object A, glimpsed a few moments ago lies relative to an object B encountered later. We introduce Disjoint-3DQA , a generative QA benchmark that evaluates this ability of VLMs by posing questions about object pairs that are not co-visible in the same frame. We evaluated seven state-of-the-art VLMs and found that models lag behind human performance by 28%, with steeper declines in accuracy (60% → 30 %) as the temporal gap widens. Our analysis further reveals that providing trajectories or bird's-eye-view projections to VLMs results in only marginal improvements, whereas providing oracle 3D coordinates leads to a substantial 20% performance increase. This highlights a core bottleneck of multi-frame VLMs in constructing and maintaining 3D scene representations over time from visual signals. Disjoint-3DQA therefore sets a clear, measurable challenge for long-horizon spatial reasoning and aims to catalyze future research at the intersection of vision, language, and embodied AI.

Out of Sight, Not Out of Context? Egocentric Spatial Reasoning in VLMs Across Disjoint Frames

Negotiation requires dynamically balancing self-interest and cooperation within the flow of conversation to maximize one’s own utility. Yet, existing agents struggle due to bounded rationality in human data, low adaptability to counterpart behavior, and limited strategic reasoning. To address this, we introduce principle-driven negotiation agents, powered by ASTRA, a novel framework for turn-level offer optimization grounded in two core principles: opponent modeling and Tit-for-Tat reciprocity. ASTRA operates in three stages: (1) interpreting counterpart behavior, (2) optimizing counteroffers via a linear programming (LP) solver, and (3) selecting offers based on negotiation tactics and the partner’s acceptance probability. Through simulations and human evaluations, our agent effectively adapts to an opponent’s shifting stance and achieves favorable outcomes through enhanced adaptability and strategic reasoning. Beyond improving negotiation performance, it also serves as a powerful coaching tool, offering interpretable strategic feedback and optimal offer recommendations.

ASTRA: A Negotiation Agent with Adaptive and Strategic Reasoning via Tool-integrated Action for Dynamic Offer Optimization

Large language models (LLMs) excel at factual recall yet still propagate stale or incorrect knowledge. In‑context knowledge editing offers a gradient-free remedy suitable for black-box APIs, but current editors rely on static demonstration sets chosen by surface-level similarity, leading to two persistent obstacles: (i) a quantity–quality trade-off, and (ii) lack of adaptivity to task difficulty. We address these issues by dynamically selecting supporting demonstrations according to their utility for the edit. We propose **D**ynamic **R**etriever for **I**n-Context **K**nowledge **E**diting (DR-IKE), a lightweight framework that (1) trains a BERT retriever with REINFORCE to rank demonstrations by editing reward, and (2) employs a *learnable threshold σ* to prune low-value examples, shortening the prompt when the edit is easy and expanding it when the task is hard. DR-IKE performs editing without modifying model weights, relying solely on forward passes for compatibility with black-box LLMs. On the CounterFact benchmark, it improves edit success by up to 17.1%, reduces latency by 41.6%, and preserves accuracy on unrelated queries—demonstrating scalable and adaptive knowledge editing.

Dynamic Retriever for In-Context Knowledge Editing via Policy Optimization

As large language models (LLMs) integrate into society, their ability to adapt to users is as critical as their accuracy. While prior work has used personality tests to examine the perceived personalities of LLMs, little research has explored whether LLMs adapt their perceived personalities in response to user interactions. We investigate whether and how LLMs exhibit conversational adaptations over prolonged interactions. Using a controlled simulations where a user and chatbot engage in dialogue, we measure the chatbot’s personality shift before and after the conversation. Across multiple models, we find that traits such as Agreeableness, Extraversion, and Conscientiousness are highly susceptible to user influence, whereas Emotional Stability and Intellect remain relatively more stable. Our results suggest that LLMs dynamically adjust their conversational style in response to user personas, raising important implications for AI alignment, trust, and safety.

Chameleon LLMs: User Personas Influence Chatbot Personality Shifts

We propose RaDeR, a set of reasoning-based dense retrieval models trained with data derived from mathematical problem solving using large language models (LLMs). Our method leverages retrieval-augmented reasoning trajectories of an LLM and self-reflective relevance evaluation, enabling the creation of both diverse and hard-negative samples for reasoning-intensive relevance. RaDeR retrievers, trained for mathematical reasoning, effectively generalize to diverse reasoning tasks in the BRIGHT and RAR-b benchmarks, consistently outperforming strong baselines in overall performance. Notably, RaDeR achieves significantly higher performance than baselines on the Math and Coding splits. In addition, RaDeR presents the first dense retriever that outperforms BM25 when queries are Chain-of-Thought reasoning steps, underscoring the critical role of reasoning-based retrieval to augment reasoning language models. Furthermore RaDeR achieves comparable or superior performance while using only 2.5% of the training data used by the concurrent work ReasonIR, highlighting the quality of our synthesized training data. Our code, data, and retrieval models are publicly available.

RaDeR: Reasoning-aware Dense Retrieval Models

Large Language Models (LLMs) are currently pre-trained and fine-tuned on large cloud servers. The next frontier is LLM personalization, where a foundation model can be fine-tuned with user/task-specific data. Given the sensitive nature of such private data, it is desirable to fine-tune these models on edge devices to improve user trust. However, fine-tuning on resource-constrained edge devices presents significant challenges due to substantial memory and computational demands, as well as limited infrastructure support. We observe that inference engines (e.g., ExecuTorch) can be repurposed for fine-tuning by leveraging zeroth-order (ZO) optimization, which uses multiple forward passes to approximate gradients. While promising, direct application of ZO methods on edge devices is inefficient due to the high computational cost of multiple forward passes required for accurate gradient estimation, and their deployment has been largely unexplored in practice. We introduce MobiZO, a resource-efficient fine-tuning framework for LLMs specifically designed for edge devices. MobiZO combines three key innovations: (1) a parallelized randomized gradient estimator that employs both outer-loop and inner-loop parallelism to eliminate sequential forward passes, (2) a specialized Multi-Perturbed LoRA (MP-LoRA) module that enables efficient realization of both inner and outer loop parallelism, and (3) a seamless integration with ExecuTorch for on-device training, requiring no modifications to the runtime. Experiments demonstrate that MobiZO achieves substantial runtime speedups and memory savings while improving fine-tuning accuracy, paving the way for practical deployment of LLMs in real-time, on-device applications. Code available at: \url{anonymous.4open.science/r/MobiZO-DBC6}.

MobiZO: Enabling Efficient LLM Fine-Tuning at the Edge via Inference Engines

Selecting an optimal set of exemplars is critical for good performance of in-context learning. However, prior exemplar search methods narrowly optimize for predictive accuracy, critically neglecting model calibration—a key determinant of trustworthiness and safe deployment. In this paper, we formulate exemplar selection as a multi-objective optimization problem, explicitly targeting both the maximization of predictive accuracy and the minimization of expected calibration error. We solve this problem with a sample-efficient Combinatorial Bayesian Optimization algorithm (COM-BOM) to find the Pareto-front that optimally trade-offs the two objectives of accuracy and calibration. We evaluate COM-BOM on multiple tasks from un-saturated MMLU-pro benchmark and find that COM-BOM beats or matches the baselines in jointly optimizing the two objectives, while requiring a minimal number of LLM API calls.

COM-BOM: Bayesian Exemplar Search for Efficiently Exploring the Accuracy-Calibration Pareto Frontier

The culture of the Post-Soviet states is complex, shaped by a turbulent history that continues to influence current events. In this study, we investigate the Post-Soviet cultural food knowledge of foundation models by constructing BORSch, a multi-modal dataset encompassing 1147 and 823 dishes in the Russian and Ukrainian languages, centered around the Post-Soviet region. We demonstrate that leading models struggle to correctly identify the origins of dishes from Post-Soviet nations in both text-only and multi-modal Question Answering (QA), instead over-predicting countries linked to the language the question is asked in. Through analysis of pre-training data, we show that these results can be explained by misleading dish-origin co-occurrences, along with linguistic phenomena such as Russian-Ukrainian code mixing. Finally, to move beyond QA-based assessments, we test models’ abilities to produce accurate visual descriptions of dishes. The weak correlation between this task and QA suggests that QA alone may be insufficient as an evaluation of cultural understanding.

Downloads

Next from EMNLP 2025

AIP: Subverting Retrieval-Augmented Generation via Adversarial Instructional Prompt

Similar lecture

Towards Faithful Natural Language Explanations: A Study Using Activation Patching in Large Language Models