United States

Audio Large language models (LLMs) are increasingly deployed in the real world, where they inevitably capture speech from unintended nearby bystanders, raising privacy risks that existing benchmarks and defences did not consider. We introduce SH-Bench, the first benchmark designed to evaluate selective hearing: a model&#39;s ability to attend to an intended main speaker while refusing to process or reveal information about incidental bystander speech. SH-Bench contains 3,968 multi-speaker audio mixtures, including both real-world and synthetic scenarios, paired with 77k multiple-choice questions that probe models under general and selective operating modes. In addition, we propose Selective Efficacy (SE), a novel metric capturing both multi-speaker comprehension and bystander-privacy protection. Our evaluation of state-of-the-art open-source and proprietary LLMs reveals substantial bystander privacy leakage, with strong audio understanding failing to translate into selective protection of bystander privacy. To mitigate this gap, we also present Bystander Privacy Fine-Tuning (BPFT), a novel training pipeline that teaches models to refuse bystander-related queries without degrading main-speaker comprehension. We show that BPFT yields substantial gains, achieving an absolute 47% higher bystander accuracy under selective mode and an absolute 16% higher SE compared to Gemini 2.5 Pro, which is the best audio LLM without BPFT. Together, SH-Bench and BPFT provide the first systematic framework for measuring and improving bystander privacy in audio LLMs.

ACL 2026

Protecting Bystander Privacy via Selective Hearing in Audio LLMs

### Welcome to the 2025 Understanding Interventions Annual Conference

This year’s event promises to be an enlightening experience, filled with opportunities for sharing, networking, engagement, and inspiration.

Our mission? To foster an inclusive environment where the science communities can come together to discuss, collaborate, and strategize on how to broaden participation in science careers. Whether you’re an educator, researcher, policymaker, or passionate advocate, this conference is designed to provide valuable insights and practical solutions.

Get ready for:
* **Cutting-edge Research:** Explore the latest trends and findings in interventions aimed at enhancing diversity and inclusion in science fields.
* **Networking Opportunities:** Connect with like-minded professionals, forge new partnerships, and exchange ideas that can drive meaningful change.
* **Inspiring Keynote & Speakers:** Gain insights from renowned speakers who are leading the charge in promoting diversity and equity in science.
* **Global Perspectives:** Engage with experts from around the world to gain diverse perspectives and approaches to addressing common challenges.
* **Professional Development:** Enhance your skills and knowledge through workshops, panel discussions, and interactive sessions tailored to your professional growth.
* **Seed Funding Opportunities:** Learn about funding opportunities and resources available to support initiatives aimed at increasing diversity in STEM.

To access this event page, you need to log in with the **email address you registered with**. Access credentials were sent to your email from Underline - subject line "Welcome to 2025 Understanding Interventions Annual Conference". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Thank you for visiting Understanding Interventions 2025 Annual Conference event page. In order to gain full access, please register [**here**](https://understandinginterventions.swoogo.com/ui2024/begin).

Please register!

2025 Understanding Interventions Annual Conference

Don’t miss out on this invaluable opportunity to be part of a movement towards a more inclusive and equitable future in science. This year’s event promises to be an enlightening experience, filled with opportunities for sharing, networking, engagement, and inspiration.
Mark your calendars, spread the word, and join us for the 2025 Understanding Interventions Annual Conference on February 7-8th.

Audio Large language models (LLMs) are increasingly deployed in the real world, where they inevitably capture speech from unintended nearby bystanders, raising privacy risks that existing benchmarks and defences did not consider. We introduce SH-Bench, the first benchmark designed to evaluate selective hearing: a model's ability to attend to an intended main speaker while refusing to process or reveal information about incidental bystander speech. SH-Bench contains 3,968 multi-speaker audio mixtures, including both real-world and synthetic scenarios, paired with 77k multiple-choice questions that probe models under general and selective operating modes. In addition, we propose Selective Efficacy (SE), a novel metric capturing both multi-speaker comprehension and bystander-privacy protection. Our evaluation of state-of-the-art open-source and proprietary LLMs reveals substantial bystander privacy leakage, with strong audio understanding failing to translate into selective protection of bystander privacy. To mitigate this gap, we also present Bystander Privacy Fine-Tuning (BPFT), a novel training pipeline that teaches models to refuse bystander-related queries without degrading main-speaker comprehension. We show that BPFT yields substantial gains, achieving an absolute 47% higher bystander accuracy under selective mode and an absolute 16% higher SE compared to Gemini 2.5 Pro, which is the best audio LLM without BPFT. Together, SH-Bench and BPFT provide the first systematic framework for measuring and improving bystander privacy in audio LLMs.

technical paper

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. 
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) 
<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>

To access this event page, you need to log in with the **email address you registered with**. 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.

Hallucinations remain a critical challenge for large language models (LLMs), particularly in Retrieval-Augmented Generation (RAG) settings where models may generate outputs unsupported by the provided context. To address this, we introduce TOHA, a TOpology-based HAllucination detector, which leverages a topological divergence metric to quantify the structural properties of graphs induced by attention matrices. Examining the topological divergence between prompt and response subgraphs in RAG settings reveals consistent patterns: higher divergence values in specific attention heads correlate with unfaithful outputs, independent of the dataset. Extensive experiments — including evaluations on question answering and summarization tasks — show that our approach achieves state-of-the-art or competitive results on several benchmarks while requiring minimal annotated data and computational resources. Our findings indicate that the topological structure of attention matrices provides an efficient and robust metric for assessing the correctness of LLM's responses.

Hallucination Detection in LLMs with Topological Divergence on Attention Graphs

Retrieving evidence for language model queries from knowledge graphs requires balancing broad search across the graph with multi-hop traversal to follow relational links. Similarity-based retrievers provide coverage but remain shallow, whereas traversal-based methods rely on selecting seed nodes to start exploration, which can fail when queries span multiple entities and relations. We introduce ARK: Adaptive Retriever of Knowledge, a tool-using KG retriever that gives a language model control over this breadth-depth tradeoff using a two-operation toolset: global lexical search over node descriptors and one-hop neighborhood exploration that composes into multi-hop traversal. ARK alternates between breadth-oriented discovery and depth-oriented expansion without depending on a fragile seed selection, a pre-set hop depth, or requiring retrieval training. ARK adapts tool use to queries, using global search for language-heavy queries and neighborhood exploration for relation-heavy queries.
On STaRK, ARK reaches 59.1\% average Hit@1 and 67.4 average MRR, improving average Hit@1 by up to 31.4\% and average MRR by up to 28.0\% over retrieval-based and agent-based training-free methods.
Finally, we distill ARK's tool-use trajectories from a large teacher into an 8B model via label-free imitation, improving Hit@1 by +7.0, +26.6, and +13.5 absolute points over the base 8B model on AMAZON, MAG, and PRIME datasets, respectively, while retaining up to 98.5\% of the teacher’s Hit@1 rate.

Autonomous Knowledge Graph Exploration with Adaptive Breadth-Depth Retrieval

Bias evaluation in large language models (LLMs) uses many metrics and benchmarks, but lacks a systematic way to measure agreement across bias metrics and models. As a result, improvements observed under one metric may contradict another, and model rankings may reflect benchmark-specific artifacts rather than stable bias profiles. In this work, we introduce Metric Agreement Score (MeAS) and Model Agreement Score (MoAS), which quantify cross-metric and cross-model agreement in bias rankings, respectively. We apply these measures to eight LLMs, seven bias metrics, and nine corpora. Our results reveal disagreement among both metrics and models: Contrary to expectations, we find that metrics within the same category (generation-based and probabilistic) often behave independently of each other. For instance, HONEST shows independence with toxicity metrics, and the Context Association Test shows no correlation with Language Modeling Bias metric. At the model level, DeepSeek-family models invert bias rankings relative to most others, indicating that the model family strongly shapes specific bias profiles. These findings challenge the assumption that bias mitigation is universally transferable and highlight the need for agreement-aware evaluation.

Quantifying Metric and Model Agreement in Bias Evaluation of Large Language Models

Emotional Support Conversation (ESC) plays a critical role in mental health assistance by providing accessible psychological support in real-world applications. Large Language Models (LLMs) have shown strong empathetic abilities in ESC tasks. Yet, existing methods overlook the issue of cognitive distortions in help-seekers’ expressions. As a result, current models can only provide basic emotional comfort, rather than helping help-seekers address their psychological distress at a deeper cognitive level. To address this challenge, we construct the CogBiasESC dataset, the first dataset that expands existing ESC datasets by adding labels for cognitive distortions, includes their type, intensity, and safe risk level. Furthermore, we propose the Cognitive Policy-driven Large Language Model framework (CoPoLLM) to enhance LLMs' ability to diagnose and intervene cognitive distortions in help-seekers. We also analyze the safety advantages of CoPoLLM from a theoretical perspective. Experimental results show that CoPoLLM significantly outperforms 15 state-of-the-art baselines in terms of distortion diagnosis accuracy, intervention strategy effectiveness, and safety risk control. Our source code is available at: https://github.com/Chips98/CoPoLLM-for-ACL-2026.

Cognitive Policy-Driven LLM for Diagnosis and Intervention of Cognitive Distortions in Emotional Support Conversation

Red team testing, an effective proactive method for evaluating the security of multimodal large language models (MLLMs), requires an expanding toolkit alongside the development of MLLM safeguards. We propose the Reference Attack, a powerful tool for red team testing against MLLMs. 
The Reference Attack is a reference-guided cross-modal jailbreak method that enhances existing prompt-to-image injection attacks by exploiting MLLMs’ semantic reconstruction capabilities. Our method embeds malicious prompts in non-text modalities (e.g., images, spreadsheets) and constructs recursive symbolic references in text, enabling MLLMs to gradually recover and generate harmful content through layered reference resolution.
The attack introduces a new vector that circumvents conventional content moderation by exploiting MLLMs' lack of security checks during cross-modal reference resolution. We evaluate the Reference Attack on leading MLLMs, including ChatGPT, Gemini, Claude, and the widely used open-source LLaMA model, and achieved an attack success rate of over 93\% across all tested models. Compared to state-of-the-art attacks, Reference Attack achieves higher success rates than all baselines under identical evaluation, with a maximum gain of 70.8\%. Our study reveals a critical gap in MLLM security and highlights the need for strict security auditing of cross-modal interactions in future content moderation.

Reference Attack: A New Cross-Modal Jailbreaking Attack against Multimodal Large Language Models

Retrieval-Augmented Generation (RAG) is widely used to ground large language models (LLMs) in external knowledge and improve factual accuracy. Prior work has explored iterative and self-reflective mechanisms to refine reasoning, but these approaches rely on internal model judgment and lack formally grounded, verifiable feedback. As a result, RAG systems may still produce logically inconsistent or contradictory answers in multi-step reasoning. In this paper, we propose LCR-RAG, a framework that integrates neuro-symbolic verification with reinforcement learning to explicitly optimize logical consistency. The core of our approach is a Logic-Consistency-driven Reward (LCR), which converts discrete logical signals—such as contradictions or incomplete inference chains—into a structured reward signal. This reward guides a PPO-based agent to iteratively rewrite queries and correct reasoning errors. Experiments on HotpotQA, ASQA, and TriviaQA show that LCR-RAG consistently outperforms strong RAG baselines, with ablation results indicating that the LCR mechanism is the primary source of improvement, even under noisy or conflicting retrieval conditions.

LCR-RAG: Enhancing Logical Consistency in Retrieval-Augmented Generation via Neuro-symbolic Reinforcement Learning

Large reasoning models (LRMs) achieve strong performance by externalizing explicit reasoning traces before producing the answer, yet suffer from overthinking challenge that allocates uniformly heavy computation to queries of varying difficulty. While proprietary models mitigate this via opaque routing, open-source LRMs still lack an efficient mechanism to internalize adaptive reasoning due to both expensive training cost and limited disclosure of training recipes. In response, we introduce RPO (Root-token Policy Optimization), a framework that enables LRMs to self-determine when to reason by training only the initial root token (e.g., whether to invoke the \<think\> tag) via group relative reward and group-wise advantages. By focusing on this pivotal branching point, RPO drastically reduces training overhead and VRAM usage. Across multiple model families and scales, RPO learns difficulty-aware adaptive thinking at just 
2% of the training compute of prior adaptive-reasoning methods.

Efficiently Learning To Reason or Not to Reason: Root-token Policy Optimization for Adaptive Thinking

Existing long-document question answering systems typically process texts as flat sequences or use heuristic chunking, which overlook the discourse structures that naturally guide human comprehension. We present a discourse-aware hierarchical framework that leverages rhetorical structure theory (RST) for long document question answering. Our approach converts discourse trees into sentence-level representations and employs LLM-enhanced node representations to bridge structural and semantic information. The framework involves three key innovations: language-universal discourse parsing for lengthy documents, LLM-based enhancement of discourse relation nodes, and structure-guided hierarchical retrieval. Extensive experiments on four datasets demonstrate consistent improvements over existing approaches through the incorporation of discourse structure, across multiple genres and languages. Moreover, the proposed framework exhibits strong robustness across diverse document types and linguistic settings.

Beyond Chunking: Discourse-Aware Hierarchical Retrieval for Long Document Question Answering

Parameter-efficient fine-tuning (PEFT) must balance effectiveness and efficiency: low-rank methods can be costly, while global representation edits often underfit token-level contexts. We propose **Token-Aware Representation Editing (TARE)**, a PEFT method that performs fine-grained, token-specific edits with a small additional inference overhead and minimal tuning.
After each FFN block in a transformer-like model, we adopt a lightweight selector that scores a small pool of hidden representation editors for each token, activates only the top-k editors, and mixes their element-wise scaling/bias updates. This design achieves superior performance while maintaining computational efficiency, yielding a more favorable Pareto frontier compared to state-of-the-art (SOTA) methods.
Across LLaMA-3-8B (eight knowledge reasoning and seven mathematical reasoning tasks) and RoBERTa-base/large (GLUE), TARE outperforms SOTAs (LoRA, DoRA, MiLoRA, LoReFT, and RED), achieving 86.7% (knowledge reasoning), 76.7% (mathematical reasoning), and 88.3% (GLUE) while tuning only 0.0392% of parameters using about 20 GiB peak GPU memory during training.
An implementation is available at: <https://github.com/PatriciaPulec/tare>.

TARE: Lightweight Token-Aware Representation Editing for Fine-tuning Transformer-like Models

Codec-based language models (LMs) have revolutionized text-to-speech (TTS). However, standard codecs entangle timbre and prosody, which hinders independent control in continuation-based LMs. To tackle this challenge, we propose DisCo-Speech, a zero-shot controllable TTS framework featuring a disentangled speech codec (DisCodec) and an LM-based generator. The core component DisCodec employs a two-stage design: 1) tri-factor disentanglement to separate speech into content, prosody, and timbre subspaces via parallel encoders and hybrid losses; and 2) fusion and reconstruction that merges content and prosody into unified content-prosody tokens suitable for LM prediction, while jointly optimizing reconstruction to address the disentanglement-reconstruction trade-off. This allows the LM to perform prosodic continuation from a style prompt while the decoder injects target timbre, enabling flexible zero-shot control. Experiments demonstrate that DisCo-Speech achieves competitive voice cloning and superior zero-shot prosody control. By resolving the core entanglement at the codec level, DisCo-Speech provides a robust foundation for controllable speech synthesis. Audio samples are available at: https://disco-speech.github.io/DisCo-demo/. Code and weights will be released at: https://github.com/disco-speech/DisCo-Speech upon acceptance.

Downloads

Next from ACL 2026

Hallucination Detection in LLMs with Topological Divergence on Attention Graphs

Similar lecture

Bystander Intervention Training for Geoscientists: Short and Long-term Workshop Effectiveness