United States

On-policy distillation (OPD), which samples trajectories from the student model and supervises them with a teacher at the token-level, avoids relying solely on verifiable terminal rewards and can yield better generalization than off-policy distillation. However, OPD requires expensive on-the-fly sampling of the student policy during training, which substantially increases training cost, especially for responses with long reasoning traces. Our initial analysis shows that, during OPD, training signals are stronger in the prefix of each output reasoning trace, and that even a short teacher-generated prefix can significantly help the student produce the correct answer. Motivated by these observations, we propose a simple yet effective modification of OPD: we apply the distillation objective only to prefixes of student-generated outputs and terminate each sampling early during distillation. Experiments on a suite of AI-for-Math and out-of-domain reasoning benchmarks show that on-policy prefix distillation matches the performance of full OPD in long reasoning outputs while reducing training FLOP by 2x–40x.

ACL 2026

Fast and Effective On-Policy Distillation from Reasoning Prefixes

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. <br>
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) <br><br><br>
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>
<br><br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to ACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

Registration Required

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

poster

Existing audio question answering benchmarks largely emphasize sound event classification or caption-grounded queries, often enabling models to succeed through shortcut strategies, short-duration cues, lexical priors, dataset-specific biases, or even bypassing audio via metadata and captions rather than genuine reasoning Thus, we present AUDITA (Audio Understanding from Diverse Internet Trivia Authors), a large-scale, real-world benchmark to rigorously evaluate audio reasoning beyond surface-level acoustic recognition. AUDITA comprises carefully curated, human-authored trivia questions grounded in real-world audio, designed to stress robust auditory reasoning through challenging distractors and long-range temporal dependencies, using probing queries that cannot be answered from isolated text or sound cues alone. Human average accuracy of 32.13% shows both the challenge of the task while demonstrating meaningful comprehension of the audio. In stark contrast, state-of-the- art audio question answering models perform poorly, with average accuracy below 8.86%. Beyond raw accuracy, we apply Item Response Theory (IRT) to estimate latent proficiency, question difficulty, and expose systematic deficiencies of the models and data.

AUDITA: A New Dataset to Audit Humans vs. AI Skill at Audio QA

Bangladesh's low-income population faces major barriers to affordable legal advice due to complex legal language, procedural opacity, and high costs. Existing AI legal assistants lack Bengali-language support and jurisdiction-specific adaptation, limiting their effectiveness. To address this, we developed \textsc{Mina}, a multilingual LLM-based legal assistant tailored for the Bangladeshi context. It employs multilingual embeddings and a RAG-based chain-of-tools framework for retrieval, reasoning, translation, and document generation, delivering context-aware legal drafts, citations, and plain-language explanations via an interactive chat interface. Evaluated by law faculty from leading Bangladeshi universities across all stages of the 2022 and 2023 Bangladesh Bar Council examinations, \textsc{Mina} achieved scores of 75–80% in the preliminary MCQs, written, and simulated viva voce components. These results matched or surpassed average human performance, demonstrating strong clarity, contextual understanding, and sound legal reasoning, while operating at approximately 0.1-0.6% of the cost of human lawyers. These results confirm its potential as a low-cost, multilingual AI assistant that automates key legal tasks and scales access to justice, offering a real-world details on building domain-specific, low-resource systems and addressing challenges of multilingual adaptation, efficiency, and sustainable public-service AI deployment.

Mina: A Multilingual LLM-Powered Legal Assistant Agent for Empowering Access to Justice in Bangladesh

Long-form audio understanding poses significant challenges for large audio language models (LALMs) due to the extreme length of audio sequences and the need to reason over heterogeneous acoustic cues distributed over time, such as speech content, speaker identity, emotion, and sound events. To address these challenges, we propose PlanRAG-Audio, a planning-based retrieval-augmented generation framework for scalable long-form audio understanding. Rather than having audio LALMs process entire recordings directly, PlanRAG-Audio explicitly plans which modalities and temporal spans are required for a given query, and retrieves only query-relevant information from a structured text and audio database. This retrieval planning enables effective reasoning over complex, cross-domain audio queries while substantially reducing the input length passed to the large language models. Experiments across a wide range of speech/audio retrieval demonstrate that PlanRAG-Audio improves reasoning accuracy and stabilizes performance as audio duration increases by decoupling inference cost from raw audio length.

PlanRAG-Audio: Planning and Retrieval Augmented Generation for Long-form Audio Understanding

Reinforcement learning optimization policies have traditionally relied on a single reward mechanism, most commonly a model-based reward. Such monolithic rewards often lack confidence calibration across domain-specific tasks and fail to capture diverse aspects of model responses. This approach requires extensive data annotation and reward model training, which is particularly challenging for multimodal models. In this work, we propose and provide a thorough study of hybrid reward and multi-aspect reward modeling. For accuracy and confidence calibration, we introduce a hybrid reward modeling framework that integrates complementary reward paradigms: model-based rewards, in which a learned reward model predicts scalar or vector scores, and rule-based reward, in which domain-specific heuristics provide explicit correctness signals with confidence. Beyond accuracy, we further incorporate multi-aspect rewards to enforce instruction adherence and introduce a generalized length-penalty reward to stabilize training and improve performance. Experiments demonstrate that this approach significantly enhances reasoning capabilities: our best-performing 3B model achieves an average improvement of ~9.5% across multimodal benchmarks, with a notable ~16% gain in mathematical reasoning tasks.

Beyond Monolithic Rewards: Hybrid Multi-Aspect Reward Optimization

Large Language Models (LLMs) can enhance reasoning capabilities through test-time scaling by generating multiple traces. However, the combination of lengthy reasoning traces with multiple sampling introduces substantial computation and high end-to-end latency. Prior work on accelerating this process has relied on similarity-based or confidence-based pruning, but these signals do not reliably indicate trace quality. To address these limitations, we propose **STEP**: **S**tep-level **T**race **E**valuation and **P**runing, a novel pruning framework that evaluates reasoning steps using hidden states and dynamically prunes unpromising traces during generation. We train a lightweight step scorer to estimate trace quality, and design a GPU memory-aware pruning strategy that triggers pruning as the GPU memory is saturated by KV cache to reduce end-to-end latency. Experiments across challenging reasoning benchmarks demonstrate that STEP reduces end-to-end inference latency by 45%–70% on average compared to self-consistency while also improving reasoning accuracy.

Hidden States as Early Signals: Step-level Trace Evaluation and Pruning for Efficient Test-Time Scaling

Large language models (LLMs) suffer significant performance degradation when user instructions and context are distributed over multiple conversational turns, yet multi-turn (MT) interactions dominate chat interfaces. The routine approach of appending full chat history to prompts rapidly exhausts context windows, leading to increased latency, higher computational costs, and diminishing returns as conversations extend. We introduce **MT-OSC**, a **O**ne-off **S**equential **C**ondensation framework that efficiently and automatically condenses chat history in the background without disrupting the user experience. MT-OSC employs a Condenser Agent that uses a few-shot inference-based Condenser and a lightweight Decider to selectively retain essential information, reducing token counts by up to 72% in 10-turn dialogues. Evaluated across 13 state-of-the-art LLMs and diverse multi-turn benchmarks, MT-OSC consistently narrows the multi-turn performance gap—yielding improved or preserved accuracy across datasets while remaining robust to distractors and irrelevant turns. Our results establish MT-OSC as a scalable solution for multi-turn chats, enabling richer context within constrained input spaces, reducing latency and operational cost, while balancing performance.

MT-OSC: Path for LLMs that Get Lost in Multi-Turn Conversation

Nonverbal vocalizations (NVs), such as laughter and sighs, are central to the expression of affective cues in emotional speech synthesis. However, learning diverse and contextually aligned NVs remains challenging in open settings due to limited NV data and the lack of explicit supervision. Motivated by this challenge, we propose Affectron as a framework for affective and contextually aligned NV generation. Built on a small-scale open and decoupled corpus, Affectron introduces an NV-augmented training strategy that expands the distribution of NV types and insertion locations. We further incorporate NV structural masking into a speech backbone pre-trained on purely verbal speech to enable diverse and natural NV synthesis. Experimental results demonstrate that Affectron produces more expressive and diverse NVs than baseline systems while preserving the naturalness of the verbal speech stream.

Affectron: Emotional Speech Synthesis with Affective and Contextually Aligned Nonverbal Vocalizations

Synthetic clinical data are essential for advancing AI in healthcare, given strict privacy constraints on electronic health records (EHRs), the scarcity of annotated data for rare or slowly progressing conditions, and demographic biases in observational cohorts. Large language models (LLMs) can generate fluent clinical text, but ensuring that such outputs are both clinically grounded and useful for downstream modeling remains challenging. We present DualAlign, a disease-agnostic framework for generating privacy-preserving, clinically faithful synthetic EHR narratives. DualAlign improves generation fidelity through two complementary alignment mechanisms: persona alignment, which conditions generation on patient demographics and risk factors, and symptom-trajectory alignment, which grounds narratives in empirically observed longitudinal symptom patterns. Using Alzheimer’s disease (AD) as a case study, DualAlign produces context-aware, symptom-rich sentences that more closely reflect real-world clinical documentation. Augmenting limited gold-standard data with DualAlign substantially improves AD symptom classification, outperforming both gold-only training and unconstrained synthetic baselines. Overall, DualAlign provides a generalizable approach for generating high-utility synthetic clinical text in chronic and progressive diseases, reducing annotation burden while enabling scalable and privacy-conscious clinical NLP research.

DualAlign: Generating Clinically Grounded Synthetic Data

Small Language Models (SLMs) provide computational advantages in resource-constrained environments, yet memory limitations remain a critical bottleneck for edge device deployment. A substantial portion of SLMs' memory footprint stems from vocabulary-related components, particularly embeddings and language modeling (LM) heads, due to large vocabulary sizes. Existing static vocabulary pruning, while reducing memory usage, suffers from rigid, one-size-fits-all designs that cause information loss during the prefill stage and lack flexibility. In this work, we identify two key principles underlying the vocabulary reduction challenge: the *lexical locality* principle, the observation that only a small subset of tokens is required during any single inference, and the *asymmetry in computational characteristics* between vocabulary-related components of SLM. Based on these insights, we introduce VocabTailor, a novel decoupled dynamic vocabulary selection framework that addresses memory constraints through offloading embedding and implements a hybrid static-dynamic vocabulary selection strategy for LM Head, enabling on-demand loading of vocabulary components. Comprehensive experiments across diverse downstream tasks demonstrate that **VocabTailor** achieves a reduction of up to 99% in the memory usage of vocabulary-related components with minimal or no degradation in task performance, substantially outperforming existing static vocabulary pruning. Our code is available at https://github.com/AwakenedInsects/VocabTailor.

VocabTailor: Dynamic Vocabulary Selection for Downstream Tasks in Small Language Models

Retrieval-Augmented Generation (RAG) systems are typically evaluated under static assumptions, despite being frequently corrected through user or expert feedback in deployment. Existing evaluation protocols focus on overall accuracy and fail to capture how systems adapt after feedback is introduced. We introduce feedback adaptation as a problem setting for RAG systems, which asks how effectively and how quickly corrective feedback propagates to future queries. To make this behavior measurable, we propose two evaluation axes: correction lag, which captures the delay between feedback provision and behavioral change, and post-feedback performance, which measures reliability on semantically related queries after feedback. Using these metrics, we show that training-based approaches exhibit a trade-off between delayed correction and reliable adaptation. We further propose PatchRAG, a minimal inference-time instantiation that incorporates feedback without retraining, demonstrating immediate correction and strong post-feedback generalization under the proposed evaluation. Our results highlight feedback adaptation as a previously overlooked dimension of RAG system behavior in interactive settings.

Downloads

Next from ACL 2026

AUDITA: A New Dataset to Audit Humans vs. AI Skill at Audio QA

Similar lecture

Long-Chain Reasoning Distillation via Adaptive Prefix Alignment