The Latest Papers about AI
-
A Broader Look into MLLM's Retrieval Heads
Abstract
Locating relevant evidence in long inputs is a core capability of Multimodal Large Language Models. Prior studies attribute this capability to retrieval heads, a sparse set of attention heads that transfer information from the context to the output. However, these analyses are confined to a limited set of modalities and the corresponding models. Moreover, how retrieval heads can be leveraged to enhance the model calls for further exploration. In this paper, we ask whether this mechanism expands together with the input modalities of the model. Taking audio-visual input as a newly added modality, we construct paired inputs in which the same answer is either spoken or shown on screen. With the evidence content fixed, retrieval heads remain sparse in the new modality and are shared to a large extent across modalities. Across 18 language, vision-language, and omni-modal checkpoints, omni-modal models largely preserve the retrieval heads of their backbones. Ablating these heads eliminates nearly all audio-visual retrieval. The heads shared across modalities make up fewer than 2\% of all heads, yet ablating them also degrades document understanding, spatial perception, and visual mathematics. Building on these findings, we turn retrieval heads into a target of optimization by training their attention with a ranking objective. This raises the page Recall@1 of Qwen3-VL-8B on MMDocIR by 6.03 points and improves document retrieval across three model families. Functional attention heads can thus serve both as a lens on multimodal models and as a target for improving them.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32411
PDF
-
TIE: Training-free CLIP Embedding Guidance for Compositional Text-to-Image Alignment
Abstract
Text-to-image (T2I) diffusion models have advanced notably, but still fail to satisfy prompt conditions, resulting in attribute mismatches and missing objects. These errors normally fall into two categories: concept loss, when an object or attribute (i.e., concept) goes missing, and concept confusion, when an attribute is assigned to the wrong object or multiple objects blend together. Recent studies suggest that these challenges are caused by the limited capacity of CLIP text encoder to capture fine-grained semantic details. Although several methods have been proposed, concept loss and concept confusion persist in recent T2I models when handling multi-object, multi-attribute prompts. In this paper, we conduct an embedding-level analysis to develop a training-free, inference-time solution for concept loss and concept confusion across CLIP-based T2I models. We find that (1) concepts mentioned later exhibit higher embedding entropy, reflecting less concentrated representations that are more often associated with concept loss, and (2) the first CLIP attention layer captures the strength of binding between each object and its attribute. Guided by our findings, we introduce TIE, a method that improves semantic alignment through a single text-embedding update. TIE addresses concept loss via entropy-aware singular value amplification and resolves concept confusion through interpolation-extrapolation binding based on CLIP attention scores, all in a training-free manner. Extensive experiments demonstrate that TIE enhances semantic fidelity in multi-concept scenarios with minimal sampling overhead. The project page is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32407
PDF
-
An Error Budget for Shuffled Multimodal Windows in the 3rd WEAR Challenge: Stacking, Count Constraints and Seam Matching
Abstract
The 3rd WEAR Dataset Challenge asks for an activity label for every one-second window of four unseen participants. Each window has 15 VideoMAEv2 frame embeddings, the accelerometer signal of one random limb, and the windows are shuffled. This report is an empirical analysis of one system built after the challenge deadline. Window classifiers are stacked with features averaged over per-participant graphs of video neighbours and scored successor candidates, and a mean-field style decoder with a soft count constraint per activity labels all windows of a participant together. The setting is offline and transductive, and the count constraint relies on each activity being performed once. In a participant-wise simulation with participant-level bootstrap intervals, stacking added 0.051 macro F1 over stage-1 fusion and the count constraint 0.011 (95% interval 0.006 to 0.017). A quadratic seam distance ranks the true same-limb successor first among about 720 candidates in 83.9% of cases, but adding it to the link model changed macro F1 by -0.0001 (-0.0006 to 0.0005). With the decoder settings and the mixing weight chosen inside the folds and the mislabelled spans left out of training, the system scores 0.936 in simulation (0.940 for participants who performed each activity once); the submitted configuration scores 0.938 (0.943). Per-participant scores range from 0.80 to 0.97. These results come from a non-nested stacking protocol that uses unlabelled features transductively, so they are not a fully inductive estimate. An oracle that averages the final unary scores over the true segments reaches 0.973, and 79% of the remaining errors involve the null class. On the private leaderboard, which no choice used, the selected Kaggle submission scored 0.9231 (public 0.9254, with bounds partly tuned there), and a later one with bounds chosen in simulation scored 0.9247 (public 0.9253).
Abstract (translated)
URL
https://arxiv.org/abs/2610.32398
PDF
-
Interactive image diagnosis using test-time backpropagation from the auxiliary task
Abstract
Multi-task deep learning has been developed to ensure explainability in computer-aided diagnosis (CAD) systems. These systems simultaneously predict interpretable characteristics of findings in auxiliary tasks to help radiologists assess the reliability of the diagnostic prediction. However, the CAD's opinion would not be referred to when the auxiliary task prediction deviates from the radiologist's expectation. To further leverage CAD's knowledge in the above situation, we proposed a multi-task gradient assist (MGA) method that brings interactivity to multi-task models without additional training. MGA allows radiologists to give feedback to the CAD and ask for additional diagnostic suggestions. Using two datasets of mammography and CT, we show that MGA successfully improves diagnostic accuracy upon feedback on various auxiliary tasks. Compared to an existing method using forward propagation only, MGA performs more robustly and can be combined to further improve accuracy. Our method paves a promising way for integrating artificial intelligence into image diagnosis.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32396
PDF
-
SABLE: Scalable Affine Bayesian Latent-space Exploration
Abstract
Complex objects such as engineering designs and proteins are often difficult to optimise because their design spaces exhibit high-dimensional structure and objective evaluations are expensive. Generative models provide continuous latent representations, but their latent spaces can have thousands to tens of thousands of dimensions, creating a regime where the latent dimension $D$ greatly exceeds the evaluation budget. We introduce SABLE, a Bayesian optimisation method for high-dimensional latent spaces with an isotropic distribution and limited budget. SABLE uses a Bayesian affine surrogate for prediction and acquisition, treating higher-order structure as surrogate discrepancy. After $N$ observations, the slope update has dimension at most $N-1$, yielding a decomposition of each query into an in-span component optimised by expected improvement (EI) in at most $N$ dimensions and an off-span component governed by a scalar norm. We show that the off-span norm implicitly selected by affine EI can lead to underexploration, and SABLE instead prescribes it using a finite-horizon schedule, without restricting search to a learned subspace. A recompression scheme reduces information to $\mathcal O(M)$ pseudo-observations, bounding refitting cost by $\mathcal O(M^2D+M^3)$ while keeping parameter inference, prediction, and acquisition decisions to tens of milliseconds. Across synthetic benchmarks and latent-space optimisation tasks, SABLE achieves high sample efficiency and reduces decision time by several orders of magnitude at large budgets.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31495
PDF
-
Co-Recursive Semantic Self-Extension: A Mathematical Foundation for Open-Ended Intelligence
Abstract
Artificial intelligence has advanced through scale, data, and computation, yet remains bounded by its representational framework. We ask what makes intelligence open-ended. Every inferential system selects meaningful distinctions, limiting what it can learn. Within a fixed purpose, performance reaches a ceiling set by its semantic domain. Teleological succession, in which purposes arise from earlier ones, faces a similar limit: Boolean-closed purpose-transformations exhaust their reachable semantic territory. We shift from purposes to inferential systems. Drawing on Kant and Gödel, we model cognitive states as partial evaluations over a formal language. Unresolved propositions may be undecided or undecidable. Reasoning can resolve the former but not the latter, leaving a frontier for extension. Non-trivial reductions of undecidable propositions expose undecidable intermediate premises whose reflective reification supports multiple decisive extensions. Finite trace-guided generation remains fallible, requiring meta-level selection and polycontextural exploration. Using Kolmogorov complexity, we show that no architecture with a fixed state description, transition mechanism, and initial condition can generate histories incorporating irreducible algorithmic information. Because physical systems are fixed in this sense, we examine collectives. Coupling finitely many finite agents does not remove the bound. Extension requires persistent external memory preserving semantic content and historical coordinates. Collective generative capacity depends on balancing overlap for transfer with difference for generation. We propose semantic self-extension as a design principle: intelligence should not merely resolve ambiguity within existing meanings, but transform unresolved structures into new spaces of meaning, generating new distinctions, questions, representations, and purposes.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31097
PDF
-
PASTA: Prompt-Aware Annealed Diffusion Sampling
Abstract
While Diffusion models produce high-quality images, real-world visual design requires intentionally preserving clean regions for headlines, logos, or UI elements. Existing layout-aware text-to-image methods struggle to strictly preserve reserved regions, often introducing unwanted content that complicates downstream editing. Existing approaches either manipulate attention to displace content or suppress conditioning via noise-level interventions. However, these methods sacrifice visual coherence, producing flat regions or noticeable style inconsistencies between reserved and non-reserved areas. To address this limitation, we propose a Prompt-Aware Shared-Latent Training-Free Annealed Diffusion Sampling (PASTA), a training-free framework that enforces reserved-region constraints by selectively applying semantic guidance to object-allowed regions during sampling. PASTA operates on a shared latent trajectory, where full-prompt and context-prompt guidance are fused through mask-guided sampling. An annealed schedule suppresses semantic leakage during early denoising stages while gradually restoring visual consistency in later stages. Extensive experiments demonstrate that PASTA improves reserved-region fidelity while preserving overall image realism and perceptual consistency.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32390
PDF
-
Acoustic and Lexical Signals for Native Language Identification of Language Learners
Abstract
Spoken native language (L1) identification for second language learners has traditionally relied on acoustic attributes of speech. Comparatively little attention has been paid to analysing speech transcripts for L1 traces. In this paper, we present a study of L1 identification using both speech and corresponding transcripts from a spoken learner English corpus. We compare a WavLM-based audio L1 classifier with a zero-shot GPT-4o classifier operating on transcripts and analyse their predictions across proficiency levels using a CEFR-balanced evaluation set. Our experiments show that the audio classifier outperforms the transcript-based approach overall when full transcripts are available, although their relative strengths vary with learner proficiency. Audio performs better at A1, whereas text performs better from A2 to C1 with full transcripts. Transcript-based accuracy generally increases with proficiency, while acoustic performance remains comparatively stable. Restricting transcript length substantially reduces text-based accuracy, whereas audio is less sensitive to input length and performs relatively well with limited input. These findings provide insights into L1 identification from acoustic and textual representations of learner speech and how performance varies with proficiency and the amount of available language material.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32382
PDF
-
Mitigating Saliency Collapse: Robust Saliency-Aware Long-Text Image-Text Alignment
Abstract
Long-text image-text alignment is essential for fine-grained scene understanding, yet existing methods mainly focus on extending context length or improving local alignment. We identify an overlooked failure mode, saliency collapse, where models over-rely on text that features visual prominence while underutilizing non-salient context. Consequently, retrieval performance degrades sharply when salient content is reduced, despite that the non-salient context remains informative. To address this, we propose RoSA, a saliency-aware finetuning framework that decomposes image-text pairs into complementary salient and contextual views, enabling joint learning of prominent and contextual semantics. We propose a modality-independent decomposition strategy to combine object-level visual saliency estimation with LLM-based text decomposition to bridge vision and language. We employ auxiliary objectives to explicitly align these views, ensuring robust representation with both salient and contextual features. RoSA greatly improves long-text retrieval and demonstrates strong robustness to saliency collapse across benchmarks. Moreover, our auxiliary objectives are plug-and-play, bringing consistent gains to existing tuning methods. Code and checkpoints are available at: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31005
PDF
-
UniWorld: A Unified World Model for Real-Time Visual World Simulation
Abstract
Continuous visual world simulation requires coupling evolving scene geometry with temporally persistent appearance. We present UniWorld, a unified framework that integrates depth trajectories and reference appearance within a shared video generation model. The central design maintains visual and geometric histories on a common timeline. Noise-aligned overlapping continuation places generated history at the prescribed sampling noise level and commits only newly generated latents. Rolling depth latent state preserves the corresponding geometric controls, incrementally reads newly available geometry, and keeps the conditioning fixed within each sampling window. Together, these mechanisms ensure exact preservation of latent history while bounding the active depth state by the window size. We evaluate continuous generation, paired history interventions, and long-horizon behavior across diverse scene categories. The same generator supports estimated, native, and on-demand simulator depth, while cached and incremental execution paths produce bitwise-identical outputs. Appearance changes and geometric control branches operate through the same interface without modifying committed history. The measured steady-state DiT execution path achieves real-time generation throughput, demonstrating the efficiency of UniWorld for continuous visual world simulation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32381
PDF
-
DealsFlow: An Evidence-Gated Agentic Harness for Venture Capital Deal Triage
Abstract
Large language models (LLMs) can read startup pitch decks, but letting them emit investment scores directly invites score inflation, run-to-run variance and rationales that are hard to audit. We describe DealsFlow, a human-in-the-loop agentic harness deployed at a UK early-stage fund. The LLM reads the deck, cites evidence and rates each category; code combines the ratings, caps them and only ever lowers them. Three mechanisms implement this: a stage-aware corroboration ceiling (78, 70 or 60 out of 100) on factual categories without external evidence; a sceptic call that may lower but never raise a score; and a pre-fetch pattern that retrieves and indexes all evidence before a single generation pass. We evaluate on $N=243$ SEIS/EIS investments with recorded outcomes. Because recent investments have unresolved outcomes, we use the 142 that are at least three years old. There, the full pipeline's score has Spearman $\rho=0.45$ (95% CI 0.32 to 0.57) with the total-value multiple and AUC 0.71 for identifying write-offs, against $\rho=0.38$ and AUC 0.68 for an unconstrained LLM. The gain is small and comes mostly from the sceptic call; we did not test whether its lower-only rule matters. Investments that later exceeded $2\times$ had 64% of factual categories corroborated, against 38% for write-offs (omnibus Kruskal-Wallis test over five outcome groups, $p=0.031$), and our retrieval was not date-restricted, so these figures may overstate screening skill. In a study of 20 inbound applications, analyst time per application fell from 44.7 to 18.9 minutes including verification and correction, and 84% of 240 cited claims were fully supported by their sources, but 70% of reports needed at least one correction.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32380
PDF
-
Sam-Correction: Fully Adaptive Label Noise Reduction for Medical Image Segmentation
Abstract
The quality of the teacher label has a dramatic impact on the outcome of the machine learning process. Medical image segmentation often suffers from a noisy ground truth (GT) mask due to difficulties in consistent annotation. Although a recent study has proposed a promising method to curate the noisy GT mask, this requires additional well-curated GT masks that are not always accessible without the aid of multiple medical experts. As an alternative, we propose SAM-Correction, which utilizes a foundation model to correct the noisy GT mask in a fully adaptive manner instead of manually preparing well-curated masks. In our approach, we progressively mitigate label noise via the segment anything model (SAM), where prompts are given from another segmentation network that we train concurrently. Experiments with various types of label noise derived from actual human annotations demonstrate that SAM-C achieves precise noise reduction and high-performing segmentation models trained thereby. Our method paves the way for robust model training even when clean labels are entirely unavailable.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32377
PDF
-
EigenGradCAM with AugSmooth: Enhancing Explainability and Precision in Multi-Stage Alzheimer's Disease Detection
2026-10-23 09:01:17Abstract
Alzheimer's disease (AD) is a progressive neurological disorder associated with memory loss, cognitive decline, and behavioral disturbances. Early and accurate diagnosis is essential, yet MRI-based deep learning models often lack interpretability, limiting their clinical use. This study presents an explainable deep learning approach for four-class AD staging using 2D axial MRI images. The method combines eigenvector-based Grad-CAM with gradient smoothing to improve saliency localization and heatmap stability. Principal component analysis is applied to gradient-weighted convolutional layers, while Gaussian noise is added to saliency weights before decomposition. To address class imbalance in the Moderate Dementia category, Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) generates synthetic samples to balance the four-stage training distribution. Custom convolutional neural networks are trained on the Best Alzheimer MRI Dataset, achieving 97.2% test accuracy, 97.0% balanced accuracy, and a Matthews correlation coefficient of 0.944. The area under the ROC curve ranges from 0.96 to 0.99 across classes. Generated heatmaps highlight the hippocampus and entorhinal cortex, regions associated with Alzheimer's disease biomarkers. These findings suggest that the proposed approach combines strong classification performance with improved model interpretability, although further clinical validation is needed.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32374
PDF
-
Past as State, Present as Attention: Persistent-State Blockwise Flow Matching for Long-Horizon Co-Speech Motion Generation
Abstract
Co-speech motion generation remains challenging over long horizons due to the need for temporal coherence and consistent style. Existing methods typically generate long motions segment by segment, either conditioning on a short seed of recent frames, which gradually drifts, or retaining ever-longer explicit histories, which are costly to maintain and hard to learn from limited data. We trace these limitations to a key asymmetry between past and present: past motion carries slowly varying style that should be compactly summarized, whereas the current segment exhibits fast-changing local dynamics that require fine-grained modeling. This asymmetry motivates a simple Past-As-State, Present-as-Attention principle for long-horizon co-speech motion generation. Following this principle, we propose PASPA, which summarizes past history with an inter-block persistent state and models the current segment with intra-block bidirectional self-attention. We instantiate PASPA within a blockwise flow matching framework, with multi-block supervision for effective training over variable histories, prefill-decode state caching for efficient rollout, and hybrid classifier-free guidance for enhanced history and speech conditioning. Extensive public-benchmark experiments show that PASPA achieves state-of-the-art FGD performance and preserves motion style across long-sequence generation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32373
PDF
-
Adaptive Control of Scientific Instruments with Nonstationary Responses through Multimodal Agents
2026-10-23 08:59:43Abstract
Controlling scientific instruments requires parameters to be adjusted progressively in response to feedback from successive operations. When unobserved operating conditions evolve with operational history, responses become nonstationary, and a static mapping from actions to responses becomes inadequate for decisions across shots. Shot-to-shot tuning of neutral beam injection (NBI) ion sources in magnetic confinement fusion devices is representative of this class of problems: the conditions of electrode surfaces and vacuum are difficult to observe directly, discharge outcomes vary between shots, and experiments are costly and carry operational risks. We propose a physically grounded multimodal agent with three independently auditable components. Deterministic algorithms extract symbolic physical quantities and visual evidence from multichannel waveforms; a memory with two layers organizes trajectories from the current task, experience from earlier tasks, and procedural knowledge, from which a multimodal large language model (MLLM) generates candidate actions; and an independent action feasibility check enforces action constraints before execution. In a strictly chronological replay of 455 shots from four EXL-50U platforms, physical symbols alone outperformed waveform images alone by 0.21-0.32 in macro-F1 for quality classification across all four evaluated models. Retrieved context increased the mean agreement with historical action directions by 22.4-42.7 percentage points, whereas the contribution of visual evidence varied across models. Under supervised operation on five ion sources, the system achieved a beam-extraction rate of 86.80% across 409 executed shots. These results demonstrate the engineering feasibility of the framework for supervised sequential tuning on a physical scientific instrument.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32370
PDF
-
A Regularity Condition Under Which a CPAB Transformation is a Diffeomorphism
Abstract
Continuous Piecewise-Affine Based (CPAB) transformations, introduced by Freifeld et al. (2015, 2017), are flows obtained by integrating continuous piecewise-affine (CPA) velocity fields. These expressive transformations have proven useful in applications including nonlinear temporal warping and trainable activation functions in 1D, as well as nonrigid spatial deformations in 2D and 3D. While computational costs were not an issue in 1D, they originally limited the practical expressiveness of CPAB transformations in higher dimensions, especially in 3D. Our recent work (Aflalo et al., 2026) removed this bottleneck, enabling efficient optimization with much finer tessellations and substantially greater expressiveness. There is, however, a mathematical subtlety that was overlooked in the original CPAB papers (Freifeld et al., 2015, 2017). CPAB transformations are always bi-Lipschitz homeomorphisms, implying that both the transformations and their inverses are differentiable almost everywhere. They are not, however, necessarily diffeomorphisms. This note remedies the omission by identifying sufficient regularity conditions under which a CPAB transformation is guaranteed to be a diffeomorphism. We explain how differentiability can fail and how the proposed conditions prevent this failure. In 1D and 2D, we show that these conditions can be easily enforced and, for commonly used tessellations, hold with probability one under standard sampling assumptions. We expect analogous behavior in 3D.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32369
PDF
-
Graphs of Trajectory Bundles for Nonlinear Motion Planning
Abstract
We present the Graphs of Trajectory Bundles (GTB) method for motion planning of systems with nonlinear and black-box dynamics, combining sampling-based motion planning with trajectory optimization and graph search. Based on a graph-of-convex-sets decomposition of the workspace, GTB iteratively constructs a graph of trajectory bundles that captures the local dynamics of the system at any vertex. By solving a shortest-path problem on this graph, we find maximally dynamically feasible iterates that satisfy the system's dynamics in the limit and any constraints such as collision avoidance or additional safety or specification constraints by construction of the underlying graph. As a result, GTB can handle highly nonlinear, unstable, and even black-box dynamics with complex collision and specification constraints, making it a powerful tool for motion planning of complex systems.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32365
PDF
-
FOCUS: A Lightweight Mamba-2 Model for EEG-Based Cognitive Workload Classification
Abstract
Cognitive workload estimation from physiological signals is increasingly relevant in high-risk environments, where continuous and unobtrusive monitoring can inform adaptive systems and safety interventions. This paper presents FOCUS (Frequency-band Optimized Cognitive-load Ultralight SSM), a compact Mamba-2-based model for EEG-based binary workload classification under constrained data conditions. Rather than learning frequency sensitivity from data, FOCUS encodes preprocessed multi-channel EEG using a fixed FFT decomposition that covers the five canonical EEG bands. The proposed model combines depthwise-separable convolution and a weight-shared encoder-bottleneck Mamba-2 pair, with 0.04 million parameters and 0.11 GFLOPs. Evaluated in a pilot study with 11 participants under leave-one-subject-out cross-validation across three cognitively distinct tasks, FOCUS achieves 74.75% accuracy on the full dataset at 0.08 ms GPU model-forward latency. In the weight-sharing ablation, the shared configuration achieves 4.55 percentage points higher accuracy on the full dataset than the non-shared variant while reducing the SSM parameter budget; however, this difference does not reach statistical significance under a two-sided Wilcoxon signed-rank test ($p=0.083$). Overall, the selected FOCUS configuration provides a favorable trade-off between classification performance and computational cost within this controlled pilot setting.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32361
PDF
-
Search-Based Testing of Modality Sensitivity in Vision Language Models
Abstract
Vision-language models (VLMs) process images and text jointly, yet robustness evaluations often perturb these inputs separately or assess only predefined corruption levels. Such evaluations provide limited evidence of which modality is more susceptible to perturbations and whether smaller changes across both modalities can jointly expose failures. We present MoST, a black-box testing framework that characterizes modality sensitivity by comparing image-only, text-only, and joint perturbations. MoST uses multi-objective search to select and combine perturbations and adjust their severity, seeking prediction degradation while limiting changes to images and prompts. We evaluate MoST on four VLMs performing object localization on ImageNet and driving images. The models exhibit distinct sensitivities: textual perturbations dominate for some, whereas others are susceptible across both modalities. In several configurations, joint perturbations match or exceed the effectiveness of single-modality perturbations with smaller modifications to each input. In a human study, 60-76% of assessments across models judge the perturbed inputs sufficiently understandable to perform the task, and human localization is more accurate than that of the corresponding VLMs. These results show that evaluating modalities separately can overlook failure-inducing combinations and that modality sensitivity provides a more detailed account of VLM robustness than aggregate failure rates alone.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32360
PDF
-
Why Full-Duplex Speech Models Collapse in Long Conversations
Abstract
Full-duplex speech models listen and speak at the same time, which makes them promising for natural, open-ended conversation. But can they sustain a coherent conversation for tens of minutes? Because they generate a token every frame, even while listening, they exhaust their context windows within minutes. Yet existing benchmarks evaluate only single turns or a few conversational rounds, too short to expose the long-horizon problem. We introduce the Full-Duplex Long-Horizon Benchmark (FD-LHB), which tests five recent full-duplex models on long replayed conversations and in live conversation with an LLM partner. Every evaluated model eventually either degrades or exhausts its context. Models that keep the full history often degrade before their context runs out, while sliding-window models still fall into repetition, silence, and gibberish. For sliding-window models, we identify two contributing mechanisms. Gibberish and silence are linked to Initialization-Anchor Eviction, where the sliding window discards the first KV entry. Repetition is also reinforced by the model's own recent outputs. Guided by this diagnosis, we apply two simple fixes: Initialization-Anchor Pinning (IAP) keeps that first entry, and Duplex Repetition Penalty (DRP) discourages echoing while guarding against prolonged silence. Together, they support the diagnosis and, without retraining, substantially reduce conversational collapse in replays of up to an hour while running in real time.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32349
PDF
