AI conference papers

AI conference paper library

Search papers from 30+ top conferences, including NeurIPS, ICLR, ICML, CVPR, ACL and AAAI. Filter by keyword, then open any paper for its abstract, authors and links.

arXiv

Cambrian-P: Pose-Grounded Video Understanding

Jihan Yang, Zifan Zhao, Xichen Pan, Shusheng Yang, Junyi Zhang, Bingyi Kang, Hu Xu, Shang-Wen Li, Saining Xie

Camera pose matters. The position and orientation of each viewpoint define a shared spatial coordinate frame that relates observations across video frames. Yet this signal is largely absent from multimodal LLMs (MLLMs) for video understanding, which process frames as isolated 2D snapshots, instead of the persistent scene humans perceive. We revisit pose as a lightweight supervisory signal and introduce Cambrian-P, a video MLLM augmented with per-frame learnable camera tokens and a pose regression head. With a carefully designed sampling scheme, the model achieves substantial gains of 4.5-6.5% on spatial reasoning benchmarks such as VSI-Bench, generalizes across eight additional spatial and general video QA benchmarks, and, as a byproduct, achieves state of the art streaming pose estimation on ScanNet. Surprisingly, training on pseudo-annotated poses from in-the-wild video further improves general video QA benchmarks, showing pose helps beyond spatial reasoning. Together, these results position camera pose as a fundamental signal for video models that reason about the physical world.

arXiv

Enhancing Glass Surface Reconstruction via Depth Prior for Robot Navigation

Jiamin Zheng, Jingwen Yu, Guangcheng Chen, Hong Zhang

Indoor robot navigation is often compromised by glass surfaces, which severely corrupt depth sensor measurements. While foundation models like Depth Anything 3 provide excellent geometric priors, they lack an absolute metric scale. We propose a training-free framework that leverages depth foundation models as a structural prior, employing a robust local RANSAC-based alignment to fuse it with raw sensor depth. This naturally avoids contamination from erroneous glass measurements and recovers an accurate metric scale. Furthermore, we introduce \ti{GlassRecon}, a novel RGB-D dataset with geometrically derived ground truth for glass regions. Extensive experiments demonstrate that our approach consistently outperforms state-of-the-art baselines, especially under severe sensor depth corruption. The dataset and related code will be released at https://github.com/jarvisyjw/GlassRecon.

arXiv

Task Alignment: A Simple Proxy for Practical Model Merging Across Diverse Vision Tasks

Pau de Jorge, César Roberto de Souza, Björn Michele, Mert Bülent Sarıyıldız, Philippe Weinzaepfel, Florent Perronnin, Diane Larlus, Yannis Kalantidis

Efficiently merging several models fine-tuned for different tasks, but stemming from the same pretrained base model, is of great practical interest. Despite extensive prior work, most evaluations of model merging in computer vision are restricted to image classification using CLIP, where different classification datasets define different tasks. In this work, our goal is to make model merging more practical and show its relevance on challenging scenarios beyond this specific setting. In most vision scenarios, different tasks rely on trainable and usually heterogeneous decoders. Differently from previous studies with frozen decoders, where merged models can be evaluated right away, the non-trivial cost of decoder training renders hyperparameter selection based on downstream performance impractical. To address this, we introduce the task alignment proxy, and show how it can be used to speed up hyperparameter selection by orders of magnitude while retaining performance. Equipped with the task alignment proxy, we extend the applicability of model merging to multi-task vision models beyond CLIP-based classification. Project page: https://europe.naverlabs.com/task-alignment

arXiv

Real-Time Human Reconstruction and Animation using Feed-Forward Gaussian Splatting

Devdoot Chatterjee, Zakaria Laskar, C. V. Jawahar

We present HumanGS, a generalizable feed-forward Gaussian splatting framework for human 3D reconstruction and real-time animation from sparse multi-view RGB images and their associated SMPL-X poses. Unlike prior methods that rely on depth supervision, fixed input views, UV maps, repeated feed-forward inference for each target pose or view, or computationally expensive vertex-to-image cross-attention, HumanGS employs a simple transformer architecture that explicitly associates SMPL-X vertices with multi-scale image features through geometric back-projection. This eliminates the need for large pre-trained human representation models while naturally aggregating complementary information from multiple views. The aggregated vertex features are mapped by a lightweight MLP decoder to a canonical set of 3D Gaussian primitives aligned with SMPL-X vertices. One Gaussian is regularized to remain close to the SMPL-X surface, providing a strong geometric prior and stable correspondence to the parametric body model, while a small set of unconstrained Gaussians per vertex captures geometric details beyond the body surface, such as clothing and hair. The resulting canonical representation is animated efficiently using linear blend skinning and Gaussian rasterization without further network inference. Trained entirely from scratch on only 10K frames from THuman2.1, HumanGS achieves reconstruction quality comparable to or better than state-of-the-art methods while reducing reconstruction time by over 15 times compared to recent transformer-based approaches, enabling real-time animation and interactive applications. Code and pre-trained models are available at https://github.com/Devdoot57/HumanGS .

arXiv

ELT: Elastic Looped Transformers for Visual Generation

Sahil Goyal, Swayam Agrawal, Gautham Govind Anil, Prateek Jain, Sujoy Paul, Aditya Kusupati

We introduce Elastic Looped Transformers (ELT), a highly parameter-efficient class of visual generative models based on a recurrent transformer architecture. While conventional generative models rely on deep stacks of unique transformer layers, our approach employs iterative, weight-shared transformer blocks to drastically reduce parameter counts while maintaining high synthesis quality. To effectively train these models for image and video generation, we propose the idea of Intra-Loop Self Distillation (ILSD), where student configurations (intermediate loops) are distilled from the teacher configuration (maximum training loops) to ensure consistency across the model's depth in a single training step. Our framework yields a family of elastic models from a single training run, enabling Any-Time inference capability with dynamic trade-offs between computational cost and generation quality, with the same parameter count. ELT significantly shifts the efficiency frontier for visual synthesis. With $4\times$ reduction in parameter count under iso-inference-compute settings, ELT achieves a competitive FID of $2.0$ on class-conditional ImageNet $256 \times 256$ and FVD of $72.8$ on class-conditional UCF-101.

arXiv

Lessons and Open Questions from a Unified Study of Camera-Trap Species Recognition Over Time

Sooyoung Jeon, Hongjie Tian, Lemeng Wang, Zheda Mai, Vidhi Bakshi, Jiacheng Hou, Ping Zhang, Arpita Chowdhury, Jianyang Gu, Wei-Lun Chao

Camera traps are vital for large-scale biodiversity monitoring, yet accurate automated analysis remains challenging due to diverse deployment environments. While the computer vision community has mostly framed this challenge as cross-domain generalization, this perspective overlooks a primary challenge faced by ecological practitioners: maintaining reliable recognition at the fixed site over time, where the dynamic nature of ecosystems introduces profound temporal shifts in both background and animal distributions. To bridge this gap, we present the first unified study of camera-trap species recognition over time. We introduce a realistic benchmark comprising 546 camera traps with a streaming protocol that evaluates models over chronologically ordered intervals. Our end-user-centric study yields four key findings. (1) Biological foundation models (e.g., BioCLIP 2) underperform at numerous sites even in initial intervals, underscoring the necessity of site-specific adaptation. (2) Adaptation is challenging under realistic evaluation: when models are updated using past data and evaluated on future intervals (mirrors real deployment lifecycles), naive adaptation can even degrade below zero-shot performance. (3) We identify two drivers of this difficulty: severe class imbalance and pronounced temporal shift in both species distribution and backgrounds between consecutive intervals. (4) We find that effective integration of model-update and post-processing techniques can largely improve accuracy, though a gap from the upper bounds remains. Finally, we highlight critical open questions, such as predicting when zero-shot models will succeed at a new site and determining whether/when model updates are necessary. Our benchmark and analysis provide actionable deployment guidelines for ecological practitioners while establishing new directions for future research in vision and machine learning.

arXiv

O3N: Omnidirectional Open-Vocabulary Occupancy Prediction for Embodied Intelligent Robotics

Mengfei Duan, Hao Shi, Fei Teng, Guoqiang Zhao, Yuheng Zhang, Zhiyong Li, Kailun Yang

The rapid evolution of consumer electronics toward embodied intelligence has accelerated the emergence of Consumer Embodied Intelligent Robotics (CEIRs), where intelligent devices are expected to perceive, understand, and interact with complex real-world environments. Understanding and reconstructing the 3D world through omnidirectional perception is therefore becoming increasingly important for CEIRs operating in complex and dynamic environments. However, existing vision-based 3D occupancy prediction methods are constrained by limited perspective inputs and a predefined training distribution, making them difficult to support embodied intelligent systems that require comprehensive and safe perception of scenes in open-world exploration. To address this, we present O3N, the first framework for open-vocabulary occupancy prediction from a single omnidirectional RGB image. O3N embeds omnidirectional voxels in a polar-spiral topology via the Polar-spiral Mamba (PsM) module, enabling continuous spatial representation and long-range context modeling across 360°. The Occupancy Cost Aggregation (OCA) module introduces a principled mechanism for unifying geometric and semantic supervision within the voxel space, ensuring consistency between reconstructed geometry and underlying semantic structure. Moreover, Natural Modality Alignment (NMA) establishes a gradient-free alignment pathway that harmonizes visual features, voxel embeddings, and text semantics, forming a consistent pixel-voxel-text representation triad for open-world perception. Extensive experiments on multiple models demonstrate that our method not only achieves state-of-the-art performance on QuadOcc and Human360Occ benchmarks but also exhibits remarkable cross-scene generalization and semantic scalability. The source code will be made publicly available at https://github.com/MengfeiD/O3N.

arXiv

QATMA: Quantization-Aware Training with Multimodal Alignment for Open-Vocabulary Object Detection

Jinyeong Park, Donghwa Kang, Seunghwan An, Insoo Kim, Brent ByungHoon Kang, Hyeongboo Baek, Jibum Kim

Quantizing open-vocabulary object detection (OVOD) models reduces their memory and computational costs, but extremely low-bit quantization severely degrades both cross-modal (region-text) and intra-modal (region-region) alignments. This multimodal degradation is a unique challenge that prior quantization methods for closed-vocabulary detectors fail to resolve. To overcome this, we propose Quantization-Aware Training with Multimodal Alignment (QATMA), the first multimodal-aware and architecture-agnostic QAT framework tailored for OVOD. QATMA integrates two key components: (i) Curriculum QAT, which partitions the detector by functional roles and progressively expands the quantization scope to suppress error accumulation and ensure stable optimization; and (ii) Text-anchored Pairwise Similarity Distillation, which transfers both region-text and region-region alignments from a full-precision teacher model via pairwise cosine similarities in the joint embedding space. Experimental results on LVIS and COCO zero-shot benchmarks demonstrate that QATMA significantly outperforms existing QAT baselines under extremely low-bit settings, achieving gains of up to 4.3 and 7.6 AP, respectively.

arXiv

Benchmarking Unlearning for Vision Transformers

Kairan Zhao, Iurie Luca, Peter Triantafillou

Machine unlearning (MU) refers to the post-training capability to remove (the influence of) training examples that are incorrect, biased, or leak sensitive/private information. MU is now widely regarded as critical for building safe and fair AI. In parallel, research into transformer architectures for computer vision has been highly successful: Vision Transformers (VTs) increasingly emerge as strong alternatives to CNNs. Yet, MU research for vision tasks has largely centered on CNNs, not VTs. While MU benchmarks have been developed for LLMs, diffusion models, and CNNs, none currently exist for VTs. This work is the first to attempt this, benchmarking MU algorithm performance across different VT families (ViT, Swin-T, and DINOv2) and at different capacities. The work employs (i) different datasets, selected to assess the impacts of dataset scale and complexity; (ii) different MU algorithms, selected to represent fundamentally different approaches for MU; and (iii) both single-shot and continual unlearning protocols. Additionally, it focuses on benchmarking MU algorithms that leverage training data memorization, since leveraging memorization has been recently discovered to significantly improve the performance of previously SOTA algorithms. En route, the work characterizes how VTs memorize training data relative to CNNs, and assesses the impact of different memorization proxies on performance. The benchmark uses unified evaluation metrics that capture two complementary notions of forget quality along with accuracy on unseen (test) data and on retained data. Overall, this work offers a benchmarking basis, enabling reproducible, fair, and comprehensive comparisons of existing (and future) MU algorithms on VTs. Importantly, for the first time, it sheds light on how well existing algorithms work in VT settings, establishing a promising reference performance baseline.

arXiv

Multimodal Learning for Arcing Detection in Pantograph-Catenary Systems

Hao Dong, Eleni Chatzi, Olga Fink

The pantograph-catenary interface is essential for ensuring uninterrupted and reliable power delivery in electrified rail systems. However, electrical arcing at this interface poses serious risks, including accelerated wear of contact components, degraded system performance, and potential service disruptions. Detecting arcing events at the pantograph-catenary interface is challenging due to their transient nature, noisy operating environment, data scarcity, and the difficulty of distinguishing arcs from other similar transient phenomena. To address these challenges, we propose a novel multimodal framework that combines high-resolution image data with force measurements to more accurately and robustly detect arcing events. First, we construct two arcing detection datasets comprising synchronized visual and force measurements. One dataset is built from data provided by the Swiss Federal Railways (SBB), and the other is derived from publicly available videos of arcing events in different railway systems and synthetic force data that mimic the characteristics observed in the real dataset. Leveraging these datasets, we propose MultiDeepSAD, an extension of the DeepSAD algorithm for multiple modalities with a new loss formulation. Additionally, we introduce tailored pseudo-anomaly generation techniques specific to each data type, such as synthetic arc-like artifacts in images and simulated force irregularities, to augment training data and improve the discriminative ability of the model. Through extensive experiments and ablation studies, we demonstrate that our framework significantly outperforms baseline approaches, exhibiting enhanced sensitivity to real arcing events even under domain shifts and limited availability of real arcing observations.

arXiv

SuperFlow: Training Flow Matching Models with RL on the Fly

Kaijie Chen, Zhiyang Xu, Ying Shen, Zihao Lin, Yuguang Yao, Lifu Huang

Recent progress in flow-based generative models and reinforcement learning (RL) has improved text-image alignment and visual quality. However, current RL training for flow models still has two main problems: (i) GRPO-style fixed per-prompt group sizes ignore variation in sampling importance across prompts, which leads to inefficient sampling and slower training; and (ii) trajectory-level advantages are reused as per-step estimates, which biases credit assignment along the flow. We propose SuperFlow, an RL training framework for flow-based models that adjusts group sizes with variance-aware sampling and computes step-level advantages in a way that is consistent with continuous-time flow dynamics. Empirically, SuperFlow reaches promising performance while using only 5.4% to 56.3% of the original training steps and reduces training time by 5.2% to 16.7% without any architectural changes. On standard text-to-image (T2I) tasks, including text rendering, compositional image generation, and human preference alignment, SuperFlow improves over SD3.5-M by 4.6% to 47.2%, and over Flow-GRPO by 1.7% to 16.0%.

arXiv

Trainable Log-linear Sparse Attention for Efficient Diffusion Transformers

Yifan Zhou, Zeqi Xiao, Tianyi Wei, Shuai Yang, Xingang Pan

Diffusion Transformers (DiTs) set the state of the art in visual generation, yet their quadratic self-attention cost fundamentally limits scaling to long token sequences. Recent Top-K sparse attention approaches reduce the computation of DiTs by compressing tokens into block-wise representation and selecting a small set of relevant key blocks, but still suffer from (i) quadratic selection cost on compressed tokens and (ii) increasing K required to maintain model quality as sequences grow. We identify that their inefficiency is due to the single-level design, as a single coarse level is insufficient to represent the global structure. In this paper, we introduce Log-linear Sparse Attention (LLSA), a trainable sparse attention mechanism for extremely long token sequences that reduces both selection and attention costs from quadratic to log-linear complexity by utilizing a hierarchical structure. LLSA performs hierarchical Top-K selection, progressively adopting sparse Top-K selection with the indices found at the previous level, and introduces a Hierarchical KV Enrichment mechanism that preserves global context while using fewer tokens of different granularity during attention computation. To support efficient training, we develop a high-performance GPU implementation that uses only sparse indices for both the forward and backward passes, eliminating the need for dense attention masks. We evaluate LLSA on high-resolution pixel-space image generation without using patchification and VAE encoding. LLSA accelerates attention inference by 28.27x and DiT training by 6.09x on 256x256 pixel token sequences, while maintaining generation quality. The results demonstrate that LLSA offers a promising direction for training long-sequence DiTs efficiently. Code is available at: https://github.com/SingleZombie/LLSA

arXiv

Focus on What Matters: Constraining Spatial-Temporal Attention via Action-Units for Noise-Resilient AQA

Shuaikang Zhu, Yiding Sun, Zihao Guo, Yang Yang, Chen Sun

The core challenge in Action Quality Assessment (AQA) lies in extracting fine-grained motion features from redundant and complex video backgrounds. Existing global feature learning methods are constrained by extremely low "signal-to-noise ratios", making it difficult to distinguish intrinsic actions from background clutter. To address this, we propose a Pose-Guided Intrinsic Motion Distillation Framework that explicitly enforces physical constraints to focus on motion subjects and decouple motion execution from environmental outcomes. First, we design an Action-Unit Parser that constructs dynamic regions of interest (ROIs) using human pose topology as prior knowledge. This functions as a spatial hard-attention filter that physically removes background noise at the input stage, forcing the model to learn appearance and geometric features only from pure body regions. Second, to resolve factor entanglement, we introduce a dual-stream decoupling mechanism: the Motion Parser focuses on capturing purified joint motion details, while the Condition Parser independently processes non-body-related environmental feedback (e.g., splash in diving) to create two orthogonal evaluation dimensions in feature space. Finally, adaptive weight modules integrate these decoupled features to generate final scores. Experimental results on large-scale datasets including FineDiving, FineDiving-HM, and MTL-AQA demonstrate that this method achieves state-of-the-art (SOTA) performance in both action segmentation and scoring accuracy, validating the effectiveness of "noise suppression focusing" and "motion disentanglement" strategies in fine-grained action evaluation.

arXiv

DCVC-MV: Deep Contextual Multiview Video Compression with Efficient Inter-View Prediction

Xihua Sheng, Yingwen Zhang, Long Xu, Shiqi Wang

Multiview video is a key format for 3D applications such as free-viewpoint broadcasting and virtual reality, yet its large data volume poses significant challenges for efficient storage and transmission. As deep contextual video compression matures and moves toward standardization, extending such learned codecs to multiview scenarios has become essential for practical deployment---yet this direction remains largely unexplored. In this paper, we propose DCVC-MV, a novel deep contextual multiview video compression framework that satisfies three fundamental requirements. First, it maintains backward compatibility, ensuring that the primary view's bitstream can be decoded independently by a single-view decoder without being affected by other views. Second, it supports random-access capability, enabling flexible switching between different viewing perspectives. Third, it effectively exploits inter-view correlations to achieve high compression efficiency. This is realized through four dedicated components: (1) an inter-view motion feature propagation method, which propagates decoded independent-view motion features as conditions to promote dependent-view motion encoding; (2) an inter-view motion conditional entropy model designed to learn motion conditional priors across views for more accurate probability estimation of motion latent representations; (3) an implicit inter-view context prediction method, which predicts inter-view contexts from low-resolution independent-view content features without explicit disparity estimation; and (4) an inter-view contextual conditional entropy model that learns contextual conditional priors across views to further enhance content compression.

arXiv

Multivariate Planar Curves: A Statistical Framework for Shape Analysis in Images

Issam-Ali Moindjié, Cédric Beaulac, Marie-Hélène Descary

Recent developments in computer vision have made segmented images widely available across many domains, such as medicine, where segmented radiographs play an important role in diagnosis. As prediction problems are common in image analysis, this work explores the use of the object contours highlighted by such images as predictors in a supervised classification context. To this end, we develop a new statistical learning framework that accounts for the joint shape of the multiple objects contained in an image. We introduce a formalism that extends the study of a single random planar curve to the joint analysis of several planar curves, referred to as a multivariate planar curve. Modeling the contours jointly, rather than separately, preserves the inter-component information, such as their relative position, scale, and orientation, which is often essential to the analysis. Based on this model, we propose a joint alignment procedure and we extend core inferential tools to multivariate shapes: shape dissimilarity, Fréchet mean estimation, and tangent-space representation. These tangent coordinates are then used as predictors in standard functional classification models. A simulation study shows accurate recovery of deformation parameters over increasing noise levels. Then, through a cardiomegaly detection problem on segmented chest X-rays, we show that jointly modeling the contours is robust to misalignment and improves classification accuracy over both a contour-wise univariate analysis and a naive approach based on the raw curves.

arXiv

SoccerSynth Field: enhancing field detection with synthetic data from virtual soccer simulator

HaoBin Qin, Jiale Fang, Keisuke Fujii

Field detection in team sports is an essential task in sports video analysis. However, collecting large-scale and diverse real-world datasets for training detection models is often cost and time-consuming. Synthetic datasets, which allow controlled variability in lighting, textures, and camera angles, will be a promising alternative for addressing these problems. This study addresses the challenges of high costs and difficulties in collecting real-world datasets by investigating the effectiveness of pretraining models using synthetic datasets. In this paper, we propose the effectiveness of using a synthetic dataset (SoccerSynth-Field) for soccer field detection. A synthetic soccer field dataset was created to pretrain models, and the performance of these models was compared with models trained on real-world datasets. The results demonstrate that models pretrained on the synthetic dataset exhibit superior performance in detecting soccer fields. This highlights the effectiveness of synthetic data in enhancing model robustness and accuracy, offering a cost-effective and scalable solution for advancing detection tasks in sports field detection.

arXiv

Evaluation and Prognostic Validation of Deep Regression Models for WSI-Based Gene-Expression Prediction

Fredrik K. Gustafsson, Constance Boissin, Johan Vallon-Christersson, Mattias Rantalainen

Gene-expression profiling is widely used in research and central to many areas of precision oncology, but remains costly and not universally accessible. Recent advances in computational pathology enable prediction of transcriptomic profiles directly from hematoxylin and eosin (H&E)-stained whole-slide images (WSIs), although optimal modeling strategies and clinical relevance remain unclear. In this study, we systematically evaluate deep regression models for WSI-based gene-expression prediction across multiple regression formulations and pathology foundation models (PFMs), and assess whether the resulting predicted transcriptomic signals retain prognostic utility. Across four TCGA datasets, we find that direct regression using attention-based multiple instance learning together with PFM feature extractors provides a strong and computationally efficient baseline, with no consistent benefit from separately training multiple models on subsets of genes. We then externally validate the selected configuration on an independent cohort of 997 breast cancer patients, demonstrating robust generalization for clinically relevant gene sets such as PAM50. To assess clinical relevance, we further evaluate predicted gene-expression scores in two independent population-representative breast cancer cohorts comprising 4,172 patients with survival endpoints, where predicted scores retain prognostic value in both the full patient cohort and the ER+ & HER2- subgroup. Together, these results demonstrate that WSI-based gene-expression prediction can generalize across independent cohorts and recover biologically and clinically meaningful molecular structure, supporting its potential as a scalable approach for transcriptomic phenotyping and risk stratification.

arXiv

Data Quality over Capacity: Internalizing Documents into LoRA Adapters for Closed-Book QA

Joan Figuerola Hurtado

We study baking documents directly into the weights of a 4-bit Gemma-4-e4b model via LoRA, so a system can answer questions about a corpus closed-book: no retrieval and no context-window budget. Across roughly 100 training runs from single documents to a 99-document corpus, we find that once adapter capacity is adequate, training-data quality is the dominant lever on closed-book accuracy, outweighing LoRA rank, learning rate, and two alternative architectures combined; capacity itself is a hard gate below which no data intervention helps. A single curation pass (shortening gold answers to canonical 1-6 word spans and dropping trivia) moved closed-book accuracy from 57.7% to 85.7% on a 15-document corpus, a larger jump than any architectural change. We confirm a capacity trend (rank must grow with corpus size) entangled with a coupling between rank and learning rate that we initially misdiagnosed. On a 15-document slice we add a real retrieval baseline: the internalized adapter (84.2% recall) beats a BM25-RAG pipeline with a base reader (58.9%) and even a realistic gold-chunk oracle (65.6%) at lower latency. We report the full arc, including three misdiagnoses, as a case study in debugging LLM training empirically.

arXiv

LeAct: Learning to Reason from Expert Actions

Ziran Yang, Chengshuai Shi, Raj Ghugare, Benjamin Eysenbach, Karthik Narasimhan, Chi Jin

Modern reasoning models depend on reasoning data, today sourced from human annotations or distilled from stronger LLMs. However, a rich and largely untapped source of supervision lies in expert systems (e.g., game engines, classical planners, theorem provers), which routinely produce near-optimal actions across diverse domains. But these experts are silent: they commit to an action without writing down the chain of thought (CoT) behind it. Recovering that CoT as natural-language reasoning would distill expert knowledge into a student that generalizes beyond the demonstrated actions. We treat it as a latent variable and study how to recover it from the action alone. Our approach, LeAct (Learning to reason from Actions), optimizes this latent variable: the student samples candidate CoTs for each expert action, and we retain those that measurably improve its own probability of recovering the action. Across imperfect-information games at multiple scales and a simulated robotics benchmark, LeAct reaches the solver's numerical floor on small enumerable games. At larger scale, it is $5\times$ closer to the solver than the strongest expert-iteration baseline. At Flop Hold'em ($\sim 10^9$ infosets), LeAct wins head-to-head by $+60$ mbb/g, and on the robotics probe it is the only training recipe that improves on direct imitation. We present a principled framework and the result: expert systems become a categorically new source of reasoning teachers for foundation models.

arXiv

Adversarial Prompts for Acceptance Collapse in Speculative Decoding

Run Wang, Chaoyi Zhou, Xi Liu, Yi Zhu, Amir Salarpour, Pedram MohajerAnsari, Zhi-Qi Cheng, Feng Luo, Siyu Huang, Mert D. Pesé

Lossless acceleration schemes, such as speculative decoding, promise significant inference speedups by relying on dynamic token-level alignment between a draft and a target model. However, this guarantee of semantic equivalence masks a severe operational vulnerability: draft-target alignment can be systematically attacked. In this paper, we introduce ADSD, which, to the best of our knowledge, is the first prompt-suffix attack that collapses verifier acceptance by pushing draft probability mass toward tokens the target is unlikely to accept. ADSD uses Soft-Collapse, a verifier-aligned surrogate derived from the asymmetric speculative acceptance rule, together with a target-preservation objective that discourages obvious task corruption. ADSD successfully generates highly effective adversarial suffixes. On the GSM8K dataset, our attack increases the mean sample time by 62.3% while preserving the task quality. We further show that this vulnerability exists across different domains, speculative decoding strategies, and model architectures.

Keep browsing

More papers from top AI conferences

30+ top conferences, by conference, year, and topic.