ACL 2026 Papers

The inability to filter out in advance all potentially problematic data from the pre-training of large language models has given rise to the need for methods for unlearning specific pieces of knowledge after training. Existing techniques overlook the need for continuous and immediate action, causing them to suffer from degraded utility as updates accumulate and protracted exposure of sensitive information. To address these issues, we propose Continual Unlearning in Real Time with Ensured Preservation of LLM Knowledge (CURaTE). Our method begins by training a sentence embedding model on a dataset designed to enable the formation of sharp decision boundaries for determining whether a given input prompt corresponds to any stored forget requests. The similarity of a given input to the forget requests is then used to determine whether to answer or return a refusal response. We show that even with such a simple approach, not only does CURaTE achieve more effective forgetting than existing methods, but by avoiding modification of the language model parameters, it also maintains near perfect knowledge preservation over any number of updates and is the only method capable of continual unlearning in real-time.

Large Language Models (LLMs) encode substantial knowledge in their parameters, which can be located, traced, and analyzed. Despite recent progress in neural interpretability, it is still unclear how to transfer such knowledge in a fine-grained manner, namely parametric knowledge transfer (PKT). A central challenge is to make cross-scale transfer effective and efficient when source and target models differ in architecture and parameterization. Existing methods that directly reuse layer parameters are therefore strongly limited by neural incompatibility. In this paper, we identify latent semantic alignment as the key prerequisite for cross-scale knowledge transfer. Instead of directly moving layer parameters, our approach uses activations as the transfer medium. SemAlign has two stages: an layer attribution stage that attributes task-relevant source layers and selects exactly one source layer for each target layer, and a semantic alignment stage that pairs them from shallow to deep and optimizes the target with source-side supervisory hidden states. The alignment is carried out in latent space. In the current realization, training follows a shallow-to-deep frontier schedule: at each stage, only the current target layer is trainable, the layer objective is a Fisher-weighted quadratic surrogate on target-space aligned logits, and the final output layer keeps KL distillation. The transferred object nonetheless remains the aligned representation itself. Evaluations on four benchmarks demonstrate the efficacy of our method. Further analysis reveals the key factors that ease cross-scale knowledge transfer and provides insights into the nature of latent semantic alignment.

Generic sentences express generalizations that tolerate exceptions without explicitly communicating information about quantities. For example, the sentence Ravens are black is true even though there are albino ravens. The sentence doesn’t explicitly communicate the number or frequency of black ravens. Whether generics semantically encode information about quantities implicitly is controversial. This work takes a large-scale distributional approach to the semantic debate. It compares thousands of naturally occurring generics and quantificational sentences using language-model probabilities. It shows that language models recover many semantic facts about quantifiers. It also shows that they recover semantic facts about surface distributional differences between generics and their “quantificational counterparts”. Accordingly, and contrary to dominant views in other fields, we formulate an empirical argument to the effect that generics are not quantificational.

Argumentation skills are an essential toolkit for large language models (LLMs). These skills are crucial in various use cases, including self-reflection, debating collaboratively for diverse answers, and countering hate speech. In this paper, we create the first benchmark for a standardized evaluation of LLM-based approaches to computational argumentation, encompassing 33 datasets from previous work in unified form. Using the benchmark, we evaluate the generalizability of five LLM families across 46 computational argumentation tasks that cover mining arguments, assessing perspectives, assessing argument quality, reasoning about arguments, and generating arguments. On the benchmark, we conduct an extensive systematic analysis of the contribution of few-shot examples, reasoning steps, model size, and training skills to the performance of LLMs on the computational argumentation tasks in the benchmark.

Translation is a fundamentally value-laden process that requires the translator to make decisions and judgments that have ethical implications. However, even though large language models (LLMs) are increasingly used for translation tasks, LLMs have not been systematically examined for their default ethical tendencies or their abilities to employ and prioritize specified ethical approaches in conflicted translation situations. To address this gap, we present ETHICA-MT, a framework for examining ethical reasoning and implementation in LLM-based machine translation. Drawing on diverse ethical approaches from the translation studies literature, we formalize a conceptual framework and construct a multilingual benchmark, ETHICA-MT BENCH, that covers six languages and highlights ethical conflicts arising from competing ethical approaches in a variety of translation scenarios. Our empirical study shows that current models predominantly default to an ethical stance favoring ‘faithful representation’ to the source text, and vary in their ability to implement specified ethics at the expense of others. Finally, we highlight the basic challenges of automatically and manually evaluating the models’ ethical stances.

Generative retrieval directly decode a document identifier (i.e., docid) in response to a query, making it impossible to provide users with explanations as an answer for “why is this document retrieved?”. To address this limitation, we propose Hierarchical Category Path-Enhanced Generative Retrieval (HyPE), which enhances explainability by first generating hierarchical category paths step-by-step then decoding docid. By leveraging hierarchical category paths which progress from broader to more specific semantic categories, HyPE can provide detailed explanation for its retrieval decision. For training, HyPE constructs category paths with external high-quality semantic hierarchy, leverages LLM to select appropriate candidate paths for each document, and optimizes the generative retrieval model with path-augmented dataset. During inference, HyPE utilizes path-aware ranking strategy to aggregate diverse topic information, allowing the most relevant documents to be prioritized in the final ranked list of docids. Our extensive experiments demonstrate that HyPE not only offers a high level of explainability but also improves the retrieval performance. We provide the code and a live demo of HyPE at https://augustinlib.github.io/HyPE/

Medical multimodal large language models are increasingly deployed in high-stakes clinical settings, yet current safety evaluations largely overlook a critical failure mode: covert semantic drift that accumulates across clinically plausible multi-turn interactions. Such drift can lead models to gradually exaggerate or conceal critical medical findings without triggering explicit safety mechanisms. We propose MSIA (Medical Semantic Infiltration Attack), a framework for modeling and inducing multi-turn medical semantic jailbreaks in clinical dialogues. MSIA enables the controlled optimization of cumulative semantic drift under stealth constraints through adaptive strategy selection and closed-loop reward feedback grounded in medical evidence. Experiments on chest X-ray–based multimodal medical dialogues show that MSIA consistently outperforms existing jailbreak methods across GPT-4o, Claude, and Gemini, achieving an average attack success rate of 76.67%. These results expose substantial and previously underestimated vulnerabilities of medical LLMs in realistic multi-turn clinical interactions. Code is available here: https://github.com/HeYamo/MSIA.

Large Reasoning Models (LRMs) allocate substantial inference-time compute to Chain-of-Thought (CoT) reasoning, improving performance on mathematics, scientific QA, and tool usage. However, this introduces overthinking: LRMs often reach a correct intermediate solution, continue reasoning, and overwrite it with an incorrect answer. We first demonstrate that oracle stopping—where we inject lt;/think gt; at every sentence boundary and select the best stopping point in hindsight—improves average accuracy by 8% while reducing thinking tokens by 72%, exposing substantial overthinking. Motivated by this finding, we propose ThinkBrake, which monitors the log-probability margin between the top continuation token and lt;/think gt; at sentence boundaries, stopping reasoning when this margin narrows. ThinkBrake requires no training and achieves favorable accuracy–efficiency trade-offs across math, scientific QA, and tool usage benchmarks, reducing thinking token usage by up to 30%. Furthermore, we provide theoretical analysis showing that ThinkBrake is equivalent to test-time realignment with a reward bonus for the lt;/think gt; token.

The dominant paradigm in video retrieval relies on embedding-based full-corpus scanning, which suffers from inherent computational inefficiency and the semantic asymmetry between information-dense videos and sparse textual queries. To bridge this gap, we introduce MAVIS, a novel multi-agent framework that rethinks retrieval as cooperative reasoning rather than brute-force search. MAVIS first bridges the granularity mismatch by parsing raw videos into a Structured Semantic Library, enabling explicit attribute-level indexing. During retrieval, a planner decomposes complex user intents into atomic sub-tasks, dispatching specialized agents to independently nominate candidates. Crucially, MAVIS employs a Logic-aware Debate mechanism with a strict veto protocol, where agents collaboratively prune logical mismatches to identify a compact set of "controversial” candidates for fine-grained verification. This agentic workflow effectively bypasses the inefficiency of full-library traversal. Extensive experiments on MSR-VTT, MSVD, and ActivityNet demonstrate that MAVIS achieves competitive performance without task-specific fine-tuning, offering a scalable and interpretable alternative to traditional dual-encoder approaches.

LLM-based universal information extraction (UIE) methods often rely on additional information beyond the original training data, which increases training complexity yet often yields limited gains. To address this, we propose ProUIE, a Macro-to-Micro progressive learning approach that improves UIE without introducing any external information. ProUIE consists of three stages: (i) macro-level Complete Modeling (CM), which learns NER, RE, and EE along their intrinsic difficulty order on the full training data to build a unified extraction foundation, (ii) meso-level Streamlined Alignment (SA), which operates on sampled data with simplified target formats, streamlining and regularizing structured outputs to make them more concise and controllable, and (iii) micro-level Deep Exploration (DE), which applies GRPO with stepwise fine-grained rewards (SFR) over structural units to guide exploration and improve performance. Experiments on 36 public datasets show that ProUIE consistently improves unified extraction, outperforming strong instruction-tuned baselines on average for NER and RE while using a smaller backbone, and it further demonstrates clear gains in production-oriented information extraction.

Sequential diagnosis requires balancing diagnostic accuracy against resource costs through iterative information gathering. Existing Large Language Model (LLM) approaches exhibit a critical knowledge-reasoning gap: despite encoding extensive medical knowledge, they struggle to reason systematically under cost constraints, often resorting to excessive testing. We propose GraphDx, a knowledge-enhanced framework with two core innovations. First, we design an automated pipeline that leverages LLMs to construct Medical Diagnosis Knowledge Graphs (MDKGs) with quantized typicality, action-centric topology, and dual-objective attributes for both diagnostic relevance and cost-sensitivity. Second, we introduce three collaborative agents (Perception, Reasoning, and Decision) where the Perception and Decision Agents handle language understanding and generation, while the Reasoning Agent performs deterministic evidence scoring and cost-aware planning on the MDKG. Experiments on MedQA and MIMIC-IV across three LLM backbones (DeepSeek-V3, Kimi-k2, Llama-3.3) show that GraphDx improves diagnostic success rates from 50–68% to 79–93% while reducing test costs by 20–54%, providing a robust, economical, and interpretable solution for automated clinical diagnosis.

Long-horizon conversational agents have to manage ever-growing interaction histories that quickly exceed the finite context windows of large language models (LLMs). Existing memory frameworks provide limited support for temporally structured information across hierarchical levels, often leading to fragmented memories and unstable long-horizon personalization. We present TiMem, a temporal–hierarchical memory framework that organizes conversations through a Temporal Memory Tree (TMT), enabling systematic memory consolidation from raw conversational observations to progressively abstracted persona representations. TiMem is characterized by three core properties: (1) temporal–hierarchical organization through TMT; (2) semantic-guided consolidation that enables memory integration across hierarchical levels without fine-tuning; and (3) complexity-aware memory recall that balances precision and efficiency across queries of varying complexity. Under a consistent evaluation setup, TiMem achieves state-of-the-art accuracy on both benchmarks, reaching 75.30% on LoCoMo and 76.88% on LongMemEval-S. It outperforms all evaluated baselines while reducing the recalled memory length by 52.20% on LoCoMo. Manifold analysis indicates clear persona separation on LoCoMo and reduced dispersion on LongMemEval-S. Overall, TiMem treats temporal continuity as a first-class organizing principle for long-horizon memory in conversational agents. The code is available at https://github.com/TiMEM-AI/timem.

Current retrieval-augmented generation (RAG) methods struggle with complex multi-hop reasoning, relying on unstructured semantic matching that lacks the logical structure needed to systematically guide retrieval. We introduce Programmatic Retrieval Optimization with Generative Reasoning and Augmented Multi-queries (PROGRAM), a novel framework that elevates retrieval to structured, program-guided reasoning. PROGRAM treats retrieval as execution of specific program types, such as logical, temporal, causal, and so forth, through three stages of ’Program-Type Selection’ with dual-metric optimization, ’Iterative Active Program Pruning’ with evidence accumulation, and ’Final Answer Generation’ with reranking. Evaluated on five benchmarks including HotPotQA, 2WikiMultihopQA, ARC-Challenge, MMLU-Pro, and MedQA with various LLMs, PROGRAM achieves state-of-the-art performance with up to 24% relative improvement on HotPotQA and 13.2% on MedQA over strong baselines including FLARE, ProbTree and Self-RAG.

Safety-critical traffic reasoning requires contrastive consistency: models must detect true hazards when an accident occurs, and reliably reject plausible-but-false hypotheses under near-identical counterfactual scenes. We present CCTVBench, a Contrastive Consistency Traffic VideoQA Benchmark built on paired real accident videos and world-model-generated counterfactual counterparts, together with minimally different, mutually exclusive hypothesis questions. CCTVBench enforces a single structured decision pattern over each video question quadruple and provides actionable diagnostics that decompose failures into positive omission, positive swap, negative hallucination, and mutual-exclusivity violation, while separating video versus question consistency. Experiments across open-source and proprietary video LLMs reveal a large and persistent gap between standard per-instance QA metrics and quadruple-level contrastive consistency, with unreliable none-of-the-above rejection as a key bottleneck. Finally, we introduce C-TCD, which leverages the semantically exclusive counterpart video as the contrast input at inference time, improving both instance-level QA and contrastive consistency.

Personality detection aims to label an individual’s traits via identifying linguistic cues from his or her written text. Previous approaches typically perform a direct mapping between text and trait labels or apply static reasoning to this task.In this paper, we argue that dynamic reasoning, underpinned by psychological theory, is essential for personality trait inference. To address this, we propose PsyPath, a novel framework that models personality detection as a process of psychologically-guided self-exploration. By enabling large language models (LLMs) to dynamically generate and answer psychologically meaningful questions, our method creates a dynamic reasoning path to explore the underlying dimensions of personality traits. This mechanism not only makes the reasoning process transparent, but also helps the model understand personality nuances in a way that mirrors expert psychological reasoning.For the “guided self-exploration”, we propose a novel hybrid scoring mechanism to step-by-step evaluate the generated nodes in the reasoning paths that balances psychological coherence (black-box scoring) and model output dynamics (white-box scoring). This reasoning-based formulation inherently reflects how psychologists assess personality, as they rely on iterative, diagnostic reasoning. Experiments on two benchmark datasets demonstrate that PsyPath consistently outperforms strong baselines, yielding improvements in predictive accuracy and model interpretability.Moreover, the generated reasoning paths provide psychologically meaningful training data, significantly improving performance and psychologically grounded interpretability in downstream tasks.

Personalized glucose regulation remains a central yet unresolved challenge in precision nutrition, as postprandial glucose response varies substantially across individuals. Existing approaches based on glycemic indices fail to adequately account for such heterogeneity and lack the mechanism to dynamically adjust meals based on personal physiological feedback. In this context, recent advances in LLM-based agents offer a promising direction, as they enable context-aware reasoning and iterative refinement. Inspired by this, we propose a physio-feedback agentic loop, a unified system that integrates individualized absorption modeling with dietary intervention to regulate glucose response. Specifically, we develop a Physiology-Aware Glucose Predictor to model individualized absorption dynamics through a learnable Temporal Physiological Absorption Decay Module. We then construct a Prediction-Driven Two-Stage Meal Optimization Agent that iteratively refines real-world meals using predicted outcomes as explicit feedback. Through extensive experiments on multiple public datasets, we demonstrate that our method not only improves prediction accuracy but also effectively reduces glucose excursions. To the best of our knowledge, this paper marks the first step in integrating physiological learning with an LLM-based agent for personalized glucose regulation.

Understanding and controlling the behavior of large language models (LLMs) is an increasingly important topic in multilingual NLP. Beyond prompting or fine-tuning, \textit{language steering}, i.e., manipulating internal representations during inference, has emerged as a more efficient and interpretable technique for adapting models to a target language. Yet, no dedicated benchmarks or evaluation protocols exist to quantify the effectiveness of steering techniques. We introduce CLaS-Bench, a lightweight parallel-question benchmark for evaluating language-forcing behavior in LLMs across 32 languages, enabling systematic evaluation of multilingual steering methods. We evaluate a broad array of steering techniques, including residual-stream DiffMean interventions, probe-derived directions, language-specific neurons, PCA/LDA vectors, Sparse Autoencoders, and prompting baselines. Steering performance is measured along two axes: language control and semantic relevance, combined into a single harmonic-mean steering score. We find that across languages simple residual-based DiffMean method consistently outperforms all other methods. Moreover, a layer-wise analysis reveals that language-specific structure emerges predominantly in later layers and steering directions cluster based on language family. CLaS-Bench is the first standardized benchmark for multilingual steering, enabling both rigorous scientific analysis of language representations and practical evaluation of steering as a low-cost adaptation alternative.

Self-Consistency improves reasoning reliability through multi-sample aggregation, but incurs substantial inference cost. Adaptive self-consistency methods mitigate this issue by adjusting the sampling budget; however, they rely on count-based stopping rules that treat all responses equally, often leading to unnecessary sampling. We propose Reliability-Aware Adaptive Self-Consistency (ReASC), which addresses this limitation by reframing adaptive sampling from response counting to evidence sufficiency, leveraging response-level confidence for principled information aggregation. ReASC operates in two stages: a single-sample decision stage that resolves instances confidently answerable from a single response, and a reliability-aware accumulation stage that aggregates responses by jointly leveraging their frequency and confidence. Across five models and four datasets, ReASC consistently achieves the best accuracy-cost trade-off compared to existing baselines, yielding improved inference efficiency across model scales from 3B to 27B parameters. As a concrete example, ReASC reduces inference cost by up to 70% relative to self-consistency while preserving accuracy on GSM8K using Gemma-3-4B-it.

Large Language Models (LLMs) as judges across various scenarios such as assessing model responses is becoming an increasingly accepted paradigm. However, existing judgment approaches often rely on trained judgers using fixed preference data, which tend to overlook diverse user preferences and struggle to adapt to real-world human-AI dialogue scenarios. To address these limitations, we propose SenseJudge, a customizable judgment framework driven by human preferences and SenseBench, a diverse and challenging instruction following benchmark derived from real-world multi-turn interactions. We applied the automatic judgment framework and benchmark to two tasks: 1) LLMs as personalized judges, and 2) model ranking. We conducted extensive experiments, and the results demonstrate that the SenseJudge framework surpasses other judgment methods and models in the LLMs-as-personalized-judges task and achieves model ranking that aligns with real human sense. Additionally, we conducted analyses on position bias and consistency, alongside ablation studies, which affirmed the robustness of SenseJudge.