Chat
-
Right Words, Wrong Moment: A Clinician-Grounded Analysis of Distress in 19,930 Conversations between Young People and ChatGPT
2026-09-28 17:53:26Marx Wang, Ella Zhang, Cameron Tan, Andrea Mock, Songling Ngo, Zijing Wang, Robert Wolfe, Shirin Amouei, Rachel A. Hanebutt, Desmond C. Ong, Caroline Figueroa, Katie Davis, Anind K. Dey, Alexis HinikerAbstract
Young people increasingly turn to General-Purpose Conversational Agents (GPCAs), such as ChatGPT, in moments of distress. We examine young adults' (ages 18-25) experiences using ChatGPT. We first collected 19,930 ChatGPT conversations and survey data from 158 young adults. We then selected five example conversations reflecting user distress. Finally, we asked ten clinicians to review those five conversations. We found distressed participants reported greater emotional engagement with ChatGPT and greater behavioral change from using it than their peers. When they turned to ChatGPT in moments of acute distress, ChatGPT was quick to give overly dramatic responses and excessive action-oriented suggestions. Clinicians endorsed ChatGPT's availability and much of its wording, but identified seven process failures, such as prematurely jumping to solutions. We translated clinicians' feedback into design guidelines following three stages: 1) asking about safety, 2) de-escalating intensity to restore emotional regulation, and 3) exploring concerns without agreeing with them.
Abstract (translated)
URL
https://arxiv.org/abs/2609.35953
PDF
-
Medical Image Alignment Assessment as a Test of Generalist Visual Reasoning in Frontier Multimodal Models
2026-09-28 17:17:27Ross Callaghan, Niannu Gao, Hojjat Azadbakht, Hui ZhangAbstract
Frontier multimodal large language models (MLLMs) are increasingly positioned as general purpose visual reasoners as part of the quest for artificial general intelligence. A key test of this generality is whether they can perform novel visual judgments that humans can make reliably from visual evidence and task instructions, without task-specific parameter optimisation. We investigate this question through the task of medical image alignment assessment, where the goal is to establish whether there is anatomical correspondence between two images. Human visual assessment of image alignment is still the gold standard and most common approach; however, it requires trained operators and is impractical to scale for large datasets. We evaluate recent generations of MLLMs on two exemplar medical image alignment tasks, varying both prompting strategies and image-presentation methods. We compare against a locally fine-tuned MLLM and a task-specific CNN to examine the trade-off between frontier general purpose models and smaller models that require specific task optimisation but can be used locally. We show that are reaching an inflection point, where frontier MLLMs can now perform effective visual assessment of medical image alignment. Models released only a few months ago generalise poorly and, in some settings, perform barely above chance, whereas GPT-6 achieves over 85% across almost all scenarios tested. Fine-tuned local models can match or exceed frontier-model performance on the tasks on which they are trained, but transfer substantially less effectively to unseen settings. These findings identify medical image alignment as a useful test bed for generalist visual reasoning and suggest that frontier multimodal models are beginning to acquire capabilities that could support a common quality-control mechanism across heterogeneous medical-imaging pipelines.
Abstract (translated)
URL
https://arxiv.org/abs/2610.06896
PDF
-
Same Bytes, Different Authority: Reserved-Token Representations in Chat-Template Prompt Injection
2026-09-28 14:51:21Yan Zhan, Yunze Song, Mengkai Hou, Wanting Zhang, Shaobo Liu, Zhijun GaoAbstract
Prompt injection against LLM agents becomes much stronger when the injected instruction is wrapped in the model's own chat template. A forged template marker such as <|im_start|> can reach the model either as a single reserved control token or as a sequence of ordinary subword tokens. The two decode to exactly the same text, and because tokenization runs on the server, the defender rather than the attacker decides which one the model receives. We use this to measure how much of the injected instruction's authority comes from the reserved token's learned representation. Encoding the forged markers as subwords, with the text held fixed and a control for the extra tokens this adds, lowers attack success on the InjecAgent benchmark by 39 to 66 percentage points on three of four open-weight families, and the gap carries over to multi-turn agent tasks in AgentDojo. On Qwen3-8B the gap is 8 points, because without reserved ids the model still recognises the forged turn from its text by reasoning; suppressing the reasoning block widens the gap to 50. The authority sits in the single learned vector at the marker position: the mean of the marker's subword vectors does not reproduce it, the vector of the nearest ordinary token restores the attack on Llama-3.1, and an adaptive attacker who searches for non-reserved markers finds such embedding neighbours on three of four families. In every base and instruction-tuned pair we test, instruction tuning strengthens the model's preference for reserved markers. The standard mitigation, a tokenizer option that encodes special tokens as ordinary subwords, applies only to tokens a configuration declares special, so in 33 of 67 distinct tokenizer configurations, covering 255 of the 400 most-downloaded chat models on Hugging Face, it leaves intact the tool-protocol tokens through which agents read untrusted tool output, and the gap persists on that channel.
Abstract (translated)
URL
https://arxiv.org/abs/2609.35932
PDF
-
MMSkillRisk: Can Agents Stay Safe When Multimodal Skills Become Traps?
2026-09-28 08:07:45Lingqi Jiang, Jialuo Chen, Jianan Ma, Xinhao Deng, Xiaohu Du, Sibo Yi, Yuqi Qing, Zhenguang Liu, Qinming He, Shiwen Cui, Changhua MenAbstract
Agent skills are shareable packages of procedural instructions, tools, and examples. Multimodal skills additionally include visual references that agents retrieve and inspect during execution. Because these images guide actions, attackers can disguise malicious instructions as ordinary visual guidance within otherwise legitimate skills. Existing skill-security research primarily examines text-carried attacks or scanner detection, leaving the runtime effects of image-borne attacks insufficiently evaluated. We introduce MMSkillRisk, to our knowledge the first publicly available benchmark dedicated to end-to-end safety evaluation of image-borne attacks in multimodal skills. To instantiate this attack surface, we design Native-Context Visual Attack (NCVA), which disguises malicious instructions as native components of teaching images, such as annotations and interface labels. The accompanying this http URL provides auxiliary guidance toward relevant visual regions without explicitly stating the malicious operation. Built from 28 curated clean skills, MMSkillRisk contains 36 attack packages and 108 executable cases spanning five attack objectives, with separate checks for attack success and legitimate-task completion. Across nine model-harness configurations evaluated in isolated sandboxes, NCVA induces unauthorized operations in every configuration. Its pooled attack success rate (ASR) reaches 43.1%, exceeding the matched text-carrier baseline by 16.4 percentage points, with higher ASR in all nine configurations. Attack success and legitimate-task completion co-occur in 36.5% of cases, reaching 72.2% for GPT-5.6-sol with Codex. These results show that skill-bundled images can induce unauthorized actions even as agents complete legitimate tasks, so task success alone does not establish safe skill use. Our code and data are available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.35912
PDF
-
Learn Now, Use Next, Trust Later: Prequential Test-Time Learning for LLM Agents
2026-09-28 08:06:03Tong Zhao, Reed Li, Yuyang Hu, Yutao Zhu, Haijin Liang, Haibo Shi, Yu Lu, Zhicheng DouAbstract
Adapting large language model agents during deployment requires not only retaining past experience, but also turning new observations into timely guidance. Many test-time learning methods, however, acquire knowledge from completed episodes. Feedback from an ongoing interaction may therefore not be distilled into knowledge soon enough to help the next decision. Acquiring knowledge at the granularity of individual transitions could reduce this delay, but raises a separate challenge: a rule that is useful within one episode may not be reliable enough to guide future episodes. Waiting for validation can forfeit immediate benefits, whereas unrestricted reuse can propagate accidental or misattributed guidance. We introduce StepLearn, a nonparametric framework that separates immediate use from persistent trust. It turns informative transitions into hypotheses that can guide the next step, while requiring prospective validation before reuse across episodes. Their predicted effects are checked against subsequent observations outside the source episodes, and only sufficiently supported hypotheses become available for persistent guidance. This process updates external knowledge while keeping all model parameters fixed. Over five rounds on WebArena-Lite and ALFWorld, StepLearn achieves average success rates of 59.9% and 84.0% with GPT-5-mini, and 57.8% and 88.1% with Qwen3.5-35B-A3B, respectively. It outperforms EvoTest, the strongest baseline, by 2.2-12.7 percentage points across the four settings. Learning dynamics further shows that these gains are not restricted to the final repetition, with advantages already present on first task attempts in most settings.
Abstract (translated)
URL
https://arxiv.org/abs/2609.35911
PDF
-
Conformal Factuality Control for Multi-Hop Retrieval-Augmented Generation
2026-09-27 23:55:03Muhammad Aimal Rehman, Chi-Kuang YehAbstract
Retrieval-augmented generation (RAG) can ground large language models in external evidence, but retrieved context does not guarantee that generated claims are factually supported. This problem is especially relevant in multi-hop RAG, where retrieval and reasoning proceed through multiple dependent stages. We study whether claim-level conformal factuality control, previously developed for RAG, remains effective in this setting. We apply split-conformal claim filtering to multi-hop RAG and evaluate it on HotpotQA, Natural Questions, and TriviaQA using Llama 3.1 8B and GPT-4o-mini, together with a single-hop reference experiment. Across all six multi-hop model-dataset configurations, increasingly stringent conformal targets consistently increase the fraction of responses whose retained claims are fully supported. At the 95% target, this rate ranges from 95.80% to 97.20%, compared with 55.60%-76.03% without filtering. However, the improvement is strongly selective: only 4.41%-31.09% of generated claims are retained and 9.70%-51.40% of responses remain non-empty at the 95% target. These results show that conformal factuality extends to multi-hop RAG, while demonstrating that nominal reliability must be interpreted jointly with claim retention and abstention.
Abstract (translated)
URL
https://arxiv.org/abs/2609.38222
PDF
-
Jev in Medicine: A Benchmark Evaluation. Preliminary Results
2026-09-27 23:33:04Alfredo Madrid-Garc\'ia, Beatriz Merino-BarbanchoAbstract
Jev is a non-generative "System One" model that assigns probabilities to predefined answer options and cannot answer outside them. Its accuracy and calibration on medical question-answering and case-based diagnostic-reasoning tasks are unknown. We evaluated Jev 1.13 on four medical benchmarks: MetaMedQA, PubMedQA, DiagnosisArena-MCQ and the NEJM Case Challenges. GPT-6 Sol, with (medium) and without reasoning, was the reference. The primary outcome was top-1 accuracy; key secondary outcomes were calibration, selective prediction and recognition of unanswerable questions. All 8,469 requests returned a valid answer. Jev's accuracy was similar to that of GPT-6 Sol with medium reasoning on PubMedQA (78.4% vs 78.2%;), lower on MetaMedQA (74.8% vs 82.7%) and much lower on DiagnosisArena-MCQ (59.8% vs 82.4%;) and the NEJM cases (61.8% vs 82.4%). On MetaMedQA, Jev's probabilities were the best calibrated (expected calibration error 0.063 vs 0.146), and its answers with a probability of at least 0.9 (52.9% of questions) were 93.4% accurate, but GPT-6 Sol was as accurate when it accepted a similar proportion of questions. On DiagnosisArena-MCQ, Jev's probabilities discriminated poorly (AUROC 0.645 vs 0.768). Of the 162 questions whose correct answer was "I don't know or cannot answer", Jev chose that option for 10.5% (GPT-6 Sol, 8.6%). Median latency was 0.27-0.31 s; all 2,823 items cost USD 0.08. Jev was fast and inexpensive, and its accuracy was similar to that of a frontier LLM on research abstracts but lower on examination questions and much lower on complex diagnostic cases. Task-specific validation is required before clinical use.
Abstract (translated)
URL
https://arxiv.org/abs/2609.34024
PDF
-
Who Warmed the Archives? LLMs Overestimate Historical Warmth
Abstract
Historical archives are an under-used source for extending the instrumental climate record backward in time, and LLMs offer a way to extract the indices climatologists derive by hand. Beyond measuring how well systems extract this signal, we check whether their errors are safe to use for cross-century comparison, since a good correlation score does not rule out systematic, era-linked bias. Comparing lexical baselines, fine-tuned historical transformers, and LLM prompting on the Pfister temperature index across five centuries of German text, lexical methods beat every fine-tuned transformer we test, including one pretrained from scratch on historical German (r=-0.016). All six LLMs we test (Gemini 2.5 Flash, GPT-5-mini, DeepSeek v4 Flash, Claude Sonnet 4.6, Qwen3.7-Plus, Kimi-K2.6-Fast) show a warm bias that grows with calendar year, with the same sign in every model (slopes +0.13 to +0.34/century, p<0.01). The effect is modest in size (r-squared approx equal to 0.01 to 0.05) but consistent across six independently developed models. The best-correlated of the six, Gemini 2.5 Flash, matches the best lexical correlation (r=0.32) at double the error. An ablation stripping explicit dates and calendar-era markers from the quotes leaves this trend essentially unchanged, favoring an anachronistic present-day prior over the model correctly inferring the quote's era. Correlation alone is thus insufficient for vetting an LLM as a historical-climate-index oracle.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37499
PDF
-
Representational Simplicity and Circuit Size Dissociate in a Threshold-Dependent Way: A Controlled Test via Adversarial Training
Abstract
Sparse-autoencoder decomposability and concentrated feature attribution are increasingly treated as evidence that a model's computation is easier to reverse-engineer. Whether this representational and attributional cleanliness actually predicts a smaller or more tractable causal circuit remains an open question. We test this directly using adversarial training as a controlled instrument: it reliably reshapes internal representations, but this alone does not constitute a test of circuit size. We investigate this question through reverse-engineering complexity: the causal structure required to recover a model's behavior at a fixed level of faithfulness. To our knowledge, this is the first controlled empirical test of whether representational or attributional simplicity translates into causal simplicity at the circuit level. Starting from the same pretrained GPT-2 Small checkpoint, we apply matched standard and adversarial continual training, requiring both conditions to retain competence on indirect object identification and pass independent robustness verification before comparing mechanisms. We then compare the models along three complementary axes: sparse-autoencoder decomposability, SAE feature engagement in task attribution, and the size of faithful circuits recovered from the raw computational graph. The robust model is more SAE-decomposable and engages fewer SAE features in task attribution. Circuit size is regime-dependent: on competence-matched IOI, standard leads or ties below 85% faithfulness, but robust needs substantially fewer edges at high faithfulness (90%, 95%), a pattern established on the primary pair while representational trends generalize across a seven-point sweep and a second corpus.
Abstract (translated)
URL
https://arxiv.org/abs/2609.35890
PDF
-
Learning to Stop without Learning to Stop: Self-Supervised Confidence Training Improves Reasoning Efficiency
2026-09-25 17:59:52Parsa Hosseini, Akasha Tigalappanavara, Sumit Nawathe, Chenrui Fan, Sourya Basu, Genta Indra Winata, Anirban Das, Soheil Feizi, Nima ChitsazanAbstract
Reasoning models often generate very long reasoning traces, making inference computationally expensive. Existing approaches typically improve efficiency either through inference-time early-stopping mechanisms or by explicitly encouraging shorter reasoning during training, for example through reinforcement learning with length penalties. We show that substantial efficiency gains can instead emerge from a different kind of supervision: \textit{confidence}. Using a self-supervised procedure, we fine-tune reasoning models to predict their confidence in the answer at intermediate points along their own reasoning trajectories using only 600 training problems. Confidence is used only as a training target: the loss contains no objective for reasoning length, efficiency, or stopping. At inference, the fine-tuned models use the standard generation procedure, with no confidence elicitation or early-stopping mechanism. Despite this, self-supervised confidence fine-tuning makes reasoning more efficient, reducing generated tokens by up to 25\% at matched accuracy across Gemma, Qwen, Nemotron, and GPT-OSS models on mathematical, scientific, and coding reasoning benchmarks, with efficiency gains comparable to methods that explicitly optimize for shorter reasoning. Analysis of reasoning episodes further shows that confidence supervision largely preserves the base models' high-level reasoning composition rather than selectively suppressing particular behaviors. Our results suggest that efficient reasoning may emerge as a downstream consequence of learning metacognitive signals, without being directly optimized.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31619
PDF
-
RePlay: Retrieval-Based Voice Playback for Multi-Turn spoken dialogue
2026-09-25 17:43:21Sathvik Udupa, Naveen Kumar, Ryan FolmsbeeAbstract
Many voice interaction applications require exact control over both the content and delivery of responses, typically using pre-recorded lines. Recent full-duplex models respond with low latency but cannot guarantee exact content or reproduce a specific recorded performance, while cascaded systems can be constrained to predefined responses at the cost of additional latency. We propose RePlay, a spoken dialogue system adapted from PersonaPlex that handles multi-turn conversations by retrieving and playing pre-recorded lines. Using probing, we identify the layer and frame at which the upcoming response becomes recoverable, and use this hidden state as the retrieval query. RePlay retains only the layers up to that point and replaces text and speech generation with lightweight turn-taking and retrieval heads. In simulated multi-turn interviews, RePlay reaches a median latency of 383 ms, 3 to 7 times lower than ASR-LLM cascades of comparable dialogue quality, at the cost of lower exact-line accuracy. In a user study, participants preferred RePlay in 63% of ratings versus 12% for a fast cascade with a small LLM (p = 0.008), and showed a non-significant preference (46% vs. 21%) over a slower cascade with a stronger LLM.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31588
PDF
-
DeepEdu-v1: Efficient and Scalable Agentic LLMs for Vietnamese Education
2026-09-25 17:33:41Quang Nguyen, Hieu Nguyen, Hien Hoang, Toan Pham, Cong Tran, Nam VuAbstract
AI tutoring could markedly improve learning outcomes for students in developing regions such as Vietnam, yet the two obvious paths both fall short. Cloud assistants such as ChatGPT route sensitive student data to foreign servers---violating data-sovereignty laws such as Vietnam's Decree 53---and, pre-trained on Western-centric corpora, are not organized around the national textbook curriculum, so their knowledge of local content is unsystematic and frequently hallucinated. Self-hosting an open model keeps data on-premise but hits a two-fold wall: post-training quantization (AWQ, GPTQ) tames the static weight footprint, yet the dynamic KV cache and prefill latency of long tutoring contexts still cause out-of-memory failures and slow responses on consumer GPUs, while the model keeps hallucinating on region-specific material. We present DeepEdu-v1, an AI-tutoring system for Vietnamese education built on SCALE (Self-improving Context-Aware Learning Engine), a framework with two innovations. First, a long-context inference engine amortizes token selection from per-sub-chunk to per-cluster granularity; on long-context retrieval it issues x7.7 fewer retrieval calls than a state-of-the-art selective-attention baseline, cutting prefill latency (TTFT) by roughly 35% while matching or improving task accuracy. Second, a self-improving agentic layer continuously curates a verified playbook from past interactions instead of fine-tuning, a design intended to progressively reduce reliance on dominant-language priors as trustworthy local knowledge accumulates. In its deployed configuration, DeepEdu achieves a nearly x2 TTFT speedup over standard vLLM serving and lifts agentic accuracy from 70.0% to 79.5% on complex tasks, with the strongest per-track gains across financial-reasoning and interactive-agent benchmarks.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31568
PDF
-
TemplateCraft: Agentic Visual Template Generation
2026-09-25 16:08:00Hongjie Yu, Zhiyuan Fan, Yuzhe Zhang, Jiangcun Du, Zhicheng Gao, Yuhong Zhang, Xiaokai Zhan, Zongshi XieAbstract
The growing popularity of short videos has driven demand for one-click content creation. Visual templates turn uploaded images into personalized content with preset effects, but reusable template generation still requires substantial manual effort in asset preparation and tool orchestration. We propose TemplateCraft, a multi-agent system that converts natural-language instructions into client-executable templates through planning, material generation, effect-workflow generation, and protocol compilation. Its Planner-Evaluator loop uses execution feedback for targeted rollback, while stage-level and long-term memory support revision without parameter updates. We evaluate TemplateCraft on TemplateBench, derived from 60 real-world templates. With the same Qwen3-VL backbone, TemplateCraft raises image/video generation success rates from 56.7%/30.0% to 66.7%/50.0% over Planner-only (best-of-three) and improves template adherence and style consistency. With additional evaluation and revision, it matches or exceeds a GPT-4o Planner-only baseline on selected metrics. Persistent assets further improve cross-input style consistency.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31451
PDF
-
ViSTA: A Simple Bridge Extends Visual Alignment to Clinical Time-Series Understanding in Multimodal LLMs
2026-09-25 16:06:29Junyi Gao, Yu Shi, Pingzhao Hu, Ewen M HarrisonAbstract
Clinical prediction models estimate risk from patient measurements, while large language models support medical text understanding and question answering. Yet their language capabilities do not ensure accurate prediction from structured, high-dimensional clinical time series. Improving this ability would connect risk estimation with flexible questions about a patient's evolving condition. We introduce ViSTA, a compact adapter that incorporates irregular numerical measurements into a pretrained vision-language model's chart representations. It learns corrections to visual tokens while leaving all pretrained parameters unchanged. On MIMIC-IV, ViSTA has the highest mean scores among the compared adaptations on all four metrics for acute kidney injury and mortality prediction across models with 2-9 billion parameters. With 0.516 million trainable parameters, the 2-billion-parameter model reaches an area under the ROC curve of 0.7376 for acute kidney injury, compared with GPT-5.6 Sol's 0.7380 with text input and high reasoning effort. Training for temporal question answering yields 69.27% accuracy at 4 billion parameters with over 90% fewer trainable parameters than low-rank adaptation using charts or numerical text, at a 2.82-4.88 percentage-point accuracy gap. ViSTA extends pretrained language models to numerical prediction and temporal questions.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31448
PDF
-
PANEL: An Open-Source, Self-Hosted Web Platform for Human Evaluation of Generative Models
2026-09-25 15:21:47Matteo Spanio, Andrea Poltronieri, Mart\'{\i}n RocamoraAbstract
Human judgement is the reference measure for evaluating generative models, yet the software used to collect it lags behing the methodology. Researchers adapt listening-test frameworks designed for perceptual protocols such as MUSHRA, rely on closed commercial survey platforms, or implement single-use web applications. Live arenas such as Chatbot Arena and Music Arena rank publicly deployed systems at scale, but do not support controlled comparisons of a laboratory's own models with its own participants. We present PANEL, an open-source, self-hosted platform for such studies. A study is authored in the browser and distributed as a single link, with audio, video, image, and text stimuli, seven question types, and screening and skip logic. The platform reports per-question summaries, across-condition significance tests, pairwise win rates and Bradley--Terry scores, and supports power analysis from pilot data. Consent versioning, self-service withdrawal, retention enforcement, and audit logging support GDPR-compliant operation. Each study exports as a machine-readable specification. PANEL is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31392
PDF
-
Mutable Transcripts: Mitigating Context Pollution through Editable Conversation State
Abstract
Contemporary large language model (LLM) chat systems treat conversation history as an immutable sequence of turns that defines the model's working context. However, user intent in real interactions is not static: it evolves through correction, refinement, and shifting constraints. This mismatch between dynamic intent and static transcripts can result in context pollution, where outdated or irrelevant information persists and continues to influence subsequent responses. We introduce mutable transcripts, a new interaction paradigm that enables users to revise prior turns through natural language edit requests, allowing the conversation history itself to be updated rather than appended. This reframes the transcript from a passive record into an editable representation of conversational state. We present a working prototype that integrates transcript-level revision into a standard chat interface and evaluate its feasibility through a controlled user study (n=17) and an illustrative transcript analysis of representative interaction scenarios. Participants significantly preferred mutable transcripts over standard chat across measures of clarity, confidence, and ease of use, with reduced intent to restart conversations. Transcript analysis of representative user study conversations shows that mutable transcripts can reduce conversation length and eliminate obsolete retained context. These findings provide initial evidence that user-driven revision of conversational history can improve interaction quality and help maintain a more current representation of user intent. The source code and prototype can be accessed at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2609.31354
PDF
-
Softmax Reparameterization for Output-Head Quantization
2026-09-25 14:01:25Asim Kadav, Christian Flores, Chirag Arora, Varun Kotte, Hongbo Zheng, Lan Yan, Priya Shanmugasundaram, Tracy Holloway KingAbstract
Large vocabularies make output heads a substantial inference cost in small language models. We propose softmax reparameterization, a post-training method that selects a functionally equivalent output head before quantization. The method subtracts a scalar multiple of the vocabulary-row mean from every output row and selects the coefficient by validation KL separately for RTN, activation-weighted MSE, and full-Hessian GPTQ. This one-dimensional search includes the original head and fixed mean-centering, preserves the full-precision softmax distribution, and leaves the trained decoder unchanged; a rank-one correction handles nonlinear logit paths such as soft-capping. Across seven heads, W4 gains concentrate where baseline quantization substantially distorts predictions: on Phi-4-mini, AW-MSE KL falls from 0.936 to 0.256. The gains survive stronger GPTQ calibration and remain complementary to exact per-channel scaling and affine quantization. Across four heads and three W4 quantizers, frozen WikiText-selected coefficients also transfer to C4 and OpenWebMath, outperforming mean-centering in all 18 comparisons where the frozen coefficient differs from $1$ and matching it in the remaining six. At W2, used as a compression stress test, benefits broaden across nearly the full model--quantizer matrix. Matched residual analysis shows that improved fidelity can accompany greater logit reconstruction error while reducing the residual's Fisher-weighted cost. For shift-compatible heads, reparameterization adds no inference operation and preserves packed W4 execution: with the decoder held in BF16, quantizing the Phi output head reduces batch-one generation latency by 10.8% relative to the BF16-head baseline.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31291
PDF
-
Who Says What: Symbolic Trimodal Binding Mechanisms in Audio-Visual LLMs
2026-09-25 12:26:23Jihoo Jung, Youngjoon Jang, Joon Son ChungAbstract
Current Audio-Visual LLMs (AVLLMs) struggle with reasoning over videos featuring multi-speaker dialogues. In such videos, resolving "who says what" is crucial, which necessitates trimodal (text-audio-visual) binding. Motivated by these challenges, we systematically investigate how this trimodal binding is achieved in AVLLMs. Specifically, we identify emergent symbolic trimodal binding mechanisms in AVLLMs that utilize modality-specific symbolic variables. By encoding auditory and visual components into symbolic variables-capturing temporal utterance sequences and spatial entity coordinates, respectively-the model establishes cross-modal linking within this abstract space. Crucially, we reveal that when trimodal binding fails, the breakdown predominantly stems from misaligned audio-visual connections. To overcome this bottleneck, we introduce an audio-visual prompting method utilizing an off-the-shelf Active Speaker Detection (ASD) model. By simply overlaying visual bounding boxes on active speakers, this training-free approach yields immediate performance gains across four conversation-centric benchmarks. Moreover, lightweight fine-tuning of fewer than 300 steps on these ASD-prompted-videos extends these gains to three general AV benchmarks, suggesting the generalizability of our method.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31193
PDF
-
Do we need to answer that question? Salience and Answerability of Potential Questions in Naturalistic Dialogue
2026-09-25 11:23:45Amandine Decker (LORIA, UL, CNRS, SEMAGRAMME, GU), Maxime Amblard (SEMAGRAMME, LORIA), Ellen Breitholtz (GU)Abstract
We empirically investigate Question Under Discussion based modelling in naturalistic dialogue by studying whether the salience of generated potential questions predicts their subsequent resolution. Building on Wu et al. (2024), we construct a dataset of 7,124 questions automatically generated from utterances and preceding context from the British National Corpus, and annotated for salience and answerability. We find a robust but low positive correlation between salience and answerability in dialogue, indicating that more salient questions are more likely to be addressed. However, this effect is markedly weaker than in monologic text, suggesting that conversational structure is less predictable. We further observe that structured interactions exhibit stronger alignment between annotators than less organised dialogues.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31130
PDF
-
Externalized CPDAG Summaries Improve LLM Causal Deduction
2026-09-25 10:11:29Wentao Sun, Jo\~ao Paulo Nogueira, Dominique Verchere, Mathieu Acher, Alonso SilvaAbstract
Corr2Cause asks whether a causal claim holds in every DAG compatible with observed correlations and conditional independencies. We frame this as latent-object reasoning: the label is defined by a CPDAG query, but free-form chain-of-thought often collapses the Markov-equivalence-class problem into local pattern matching. We propose Structured Thinking, a two-turn pipeline that first externalizes a typed, schema-constrained CPDAG summary and then answers against that graph state. On the Corr2Cause full test, Structured Thinking raises Qwen3.5-27B from $73.0$ to $86.4$ $F_1$(Yes) over a strong PC-instruction baseline in the primary paired run ($+13.4$ pp; McNemar $p=2.4\times 10^{-6}$; bootstrap $95\%$ CI [$+8.4$, $+18.6$]); across three full-ID seeds, the mean gain is $+8.1 \pm 5.3$ pp. A PC-scaffolded two-turn prose control reaches only $67.6$ $F_1$, indicating that a detailed PC scaffold plus a schema-free prose intermediate is not sufficient. The same pattern holds on Qwen3.6-27B, Paraphrase-OOD, and GPT-5.4-mini. Scrambling the emitted CPDAG costs $12.0$ pp $F_1$, and a full-split audit shows close agreement with the reference CPDAG (ID skeleton $F_1$ $0.960$; exact match $75.9\%$). These results support a bounded design principle: externalize the latent object that defines the label, constrain its form, and test whether downstream answers use it.
Abstract (translated)
URL
https://arxiv.org/abs/2609.31071
PDF
