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

DuetHOI: Language-Guided Bimanual Hand--Object Motion Generation with Articulation Planning and Contact Refinement

Zhi Wang, Yuyan Liu, Liu Liu, Ruonan Liu, Ruixuan Liu

Bimanual articulated-object interaction generation requires a model to capture the evolution of object articulation, coordination between the two hands, and fine-grained hand--object contact. Existing methods typically encode object and hand motion as a unified high-dimensional sequence, making it difficult to explicitly accommodate the different scales of manipulation progress, relative bimanual motion, and local hand pose. We propose \textbf{DuetHOI}, a structured framework for bimanual articulated-object interaction generation. DuetHOI first uses ContactVAE to predict contact intent on the object surface from language and object geometry. Conditioned on this spatial intent, ArtPlanner provides a trajectory-level articulation reference, and DualFormer generates the global interaction using object states, object-centered hand positions, and compact hand-pose tokens learned by ManiVAE. After global generation, ProxiRefine fixes the object trajectory and residually corrects the two hands using current hand--object proximity, improving local surface alignment. Experiments across articulated and rigid objects and single- and bimanual settings show that DuetHOI outperforms three adapted baselines on most contact and hand--object consistency metrics, with particularly strong results in bimanual articulated interaction generation.

arXiv

Step-Level Visual Grounding Faithfulness Predicts Out-of-Distribution Generalization in Long-Horizon Vision-Language Models

Md Ashikur Rahman, Md Arifur Rahman, Niamul Hassan Samin, Abdullah Ibne Hanif Arean, Juena Ahmed Noshin

We uncover a behavioral law of long-horizon vision-language models: models that maintain temporally grounded beliefs generalize better. Standard benchmarks measure only final-answer accuracy, which obscures how models use visual information; a model can guess correctly while its step-by-step reasoning is entirely unanchored to the visual input. We formalize this as behavioral faithfulness over long horizons, an empirically measurable property that quantifies whether a model's intermediate reasoning remains consistent with the evolving visual state. Across eight models on three long-horizon benchmarks, we demonstrate that temporal grounding quality is a leading indicator of robustness: the Step Grounding Rate (SGR) predicts out-of-distribution retention with $r = 0.83$ (permutation test $p = 0.003$), a relationship that holds within capacity-matched models and cannot be explained by scale or in-distribution accuracy. Critically, grounding quality varies by up to 10.8 percentage points within parameter-matched 7B models despite similar accuracy, revealing it as an independent axis of model capability. Multiple robustness checks confirm the signal reflects genuine visual reliance: counterfactual traces drop SGR by 26--41 percentage points, cross-architecture verifiers agree at $ρ= 0.96$, random reasoning scores near chance ($\sim 18\%$), and the predictor remains strong even without explicit reasoning disclosure ($r = 0.78$).

arXiv

Prompt Reinjection: Alleviating Prompt Forgetting in Multimodal Diffusion Transformers for Text-to-Image Generation

Yuxuan Yao, Yuxuan Chen, Hui Li, Kaihui Cheng, Qipeng Guo, Yuwei Sun, Zilong Dong, Jingdong Wang, Siyu Zhu

Multimodal Diffusion Transformers (MMDiTs) for text-to-image generation maintain separate text and image branches, with bidirectional information flow between text tokens and visual latents throughout denoising. In this setting, we observe a prompt forgetting phenomenon: the semantics of the prompt representation in the text branch is progressively forgotten as depth increases. We further verify this effect on three representative MMDiTs--SD3, SD3.5, and FLUX.1 by probing linguistic attributes of the representations over the layers in the text branch. Motivated by these findings, we introduce a training-free approach, prompt reinjection, which reinjects prompt representations from early layers into later layers to alleviate this forgetting. Experiments on GenEval, DPG, and T2I-CompBench++ show consistent gains in instruction-following capability, along with improvements on metrics capturing preference, aesthetics, and overall text--image generation quality.

arXiv

The Geometric Observability Index: Influence, Fisher Information, and Weak Observability in SE(3) Pose Estimation

Joe-Mei Feng, Hsin-Hsiung Kao, Sheng-Wei Yu

We introduce the Geometric Observability Index (GOI), a per-feature sensitivity measure for pose estimation on SE(3): the metric norm of the pose perturbation that a single measurement induces through the (possibly rank-deficient) Gauss-Newton curvature, restricted to the observable subspace. We prove that GOI equals the norm of the M-estimator influence function, that the underlying curvature operator coincides with the Fisher information, and that its smallest observable eigenvalue governs both the worst-case amplification of a measurement's effect and a finite-sample stability radius O(sigma/sqrt(n*lambda_min)). Operationally the theory cuts both ways. GOI is the exact per-measurement attribution, predicting the true leave-one-out pose shift with log-correlation r = 1.00; yet the influence standardized by its inlier null covariance collapses exactly to the classical chi-square residual statistic. Residual gating is thus the leverage-corrected influence test -- a first-principles explanation of its robustness -- while raw-influence gating conflates a measurement's information with its harm and is predicted to over-reject high-leverage inliers in weakly observable geometry. Controlled synthetic experiments validate every quantitative claim, and studies on five TUM RGB-D sequences (four dynamic, one static control) and two KITTI odometry sequences confirm the prediction: parity of the two criteria under well-conditioned geometry, and significant degradation of raw-influence gating at cond(H) of order 10^4. All code is released for reproducibility.

arXiv

Progressive Checkerboards for Autoregressive Multiscale Image Generation

David Eigen

A key challenge in autoregressive image generation is to efficiently sample independent locations in parallel, while still modeling mutual dependencies with serial conditioning. Some recent works have addressed this by conditioning between scales in a multiscale pyramid. Others have looked at parallelizing samples in a single image using regular partitions or randomized orders. In this work we examine a flexible, fixed ordering based on progressive checkerboards for multiscale autoregressive image generation. Our ordering draws samples in parallel from evenly spaced regions at each scale, maintaining full balance in all levels of a quadtree subdivision at each step. This enables effective conditioning both between and within scales. Intriguingly, we find evidence that in our balanced setting, a wide range of scale-up factors lead to similar results, so long as the total number of serial steps is constant. On class-conditional ImageNet, our method achieves competitive performance compared to recent state-of-the-art autoregressive systems with like model capacity, using fewer sampling steps.

arXiv

EMAG: Self-Rectifying Diffusion Sampling with Exponential Moving Average Guidance

Ankit Yadav, Ta Duc Huy, Lingqiao Liu

In diffusion and flow-matching generative models, guidance techniques are widely used to improve sample quality and consistency. Classifier-free guidance (CFG) is the de facto choice in modern systems and achieves this by contrasting conditional and unconditional samples. Recent work explores contrasting negative samples at inference using a weaker model, via strong/weak model pairs, attention-based masking, stochastic block dropping, or perturbations to the self-attention energy landscape. While these strategies refine the generation quality, they still lack a reliable control over the granularity or difficulty of the negative samples, and target-layer selection is often fixed. We propose Exponential Moving Average Guidance (EMAG), a training-free mechanism that modifies attention at inference time in diffusion transformers, with a statistics-based, adaptive layer-selection rule. Unlike prior methods, EMAG produces harder, semantically faithful negatives (fine-grained degradations), surfacing difficult failure modes, enabling the denoiser to refine subtle artifacts, boosting the quality and human preference score (HPS) by +0.54 over CFG. We further demonstrate that EMAG naturally composes with advanced orthogonal guidance techniques, such as APG and CADS, further improving HPS.

arXiv

JoVA: Unified Multimodal Learning for Joint Video-Audio Generation and Editing

Xiaohu Huang, Haoyang He, Hao Zhou, Qiangpeng Yang, Shilei Wen, Kai Han

In this paper, we present JoVA, a streamlined framework that unifies joint video-audio generation and editing. While existing methods often rely on fragmented, task-specific architectures or complex fusion mechanisms, JoVA employs native joint representation learning for direct video, audio, and text interaction in a dual-branch architecture. This design eliminates redundant alignment modules and effectively unifies diverse multimodal tasks within a single model. Furthermore, we utilize channel-wise conditioning for flexible image and video reference to avoid massive token expansion, alongside a mouth-area loss to enhance lip alignment. To fully empower and systematically evaluate this framework, we construct a comprehensive training corpus encompassing video-audio generation and editing datasets, and introduce unified benchmarks tailored for these multimodal tasks. Extensive experiments demonstrate that JoVA achieves state-of-the-art performance across benchmarks, establishing it as an extensible framework for versatile content creation. Project page: https://visual-ai.github.io/jova

arXiv

Live Avatar: Streaming Real-time Audio-Driven Avatar Generation with Infinite Length

Yubo Huang, Hailong Guo, Fangtai Wu, Weiqiang Wang, Shifeng Zhang, Shijie Huang, Qijun Gan, Lin Liu, Sirui Zhao, Enhong Chen, Jiaming Liu, Steven Hoi

Audio-driven avatar interaction demands real-time, streaming, and infinite-length generation -- capabilities fundamentally at odds with the sequential denoising and long-horizon drift of current diffusion models. We present Live Avatar, an algorithm-system co-designed framework that addresses both challenges for a 14-billion-parameter diffusion model. On the algorithm side, a two-stage pipeline distills a pretrained bidirectional model into a causal, few-step streaming one, while a set of complementary long-horizon strategies eliminate identity drift and visual artifacts, enabling stable autoregressive generation exceeding 10000 seconds. On the system side, Timestep-forcing Pipeline Parallelism (TPP) assigns each GPU a fixed denoising timestep, converting the sequential diffusion chain into an asynchronous spatial pipeline that simultaneously boosts throughput and improves temporal consistency. Live Avatar achieves 45 FPS with a TTFF of 1.21\,s on 5 H800 GPUs, and to our knowledge is the first to enable practical real-time streaming of a 14B diffusion model for infinite-length avatar generation. We further introduce GenBench, a standardized long-form benchmark, to facilitate reproducible evaluation. Our project page is at https://liveavatar.github.io/.

arXiv

Fast Feature Field ($\text{F}^3$): A Predictive Representation of Events

Richeek Das, Kostas Daniilidis, Pratik Chaudhari

This paper develops a mathematical argument and algorithms for building representations of data from event-based cameras, that we call Fast Feature Field ($\text{F}^3$). We learn this representation by predicting future events from past events and show that it preserves scene structure and motion information. $\text{F}^3$ exploits the sparsity of event data and is robust to noise and variations in event rates. It can be computed efficiently using ideas from multi-resolution hash encoding and deep sets - achieving 120 Hz at HD and 440 Hz at VGA resolutions. $\text{F}^3$ represents events within a contiguous spatiotemporal volume as a multi-channel image, enabling a range of downstream tasks. We obtain state-of-the-art performance on optical flow estimation, semantic segmentation, and monocular metric depth estimation, on data from three robotic platforms (a car, a quadruped robot and a flying platform), across different lighting conditions (daytime, nighttime), environments (indoors, outdoors, urban, as well as off-road) and dynamic vision sensors (resolutions and event rates). Our implementations can predict these tasks at 25-75 Hz at HD resolution.

arXiv

WaMo: Wavelet-Enhanced Multi-Frequency Trajectory Analysis for Fine-Grained Text-Motion Retrieval

Junlong Ren, Gangjian Zhang, Honghao Fu, Pengcheng Wu, Hao Wang

Text-Motion Retrieval (TMR) aims to retrieve 3D motion sequences semantically relevant to text descriptions. However, matching 3D motions with text remains highly challenging, primarily due to the intricate structure of the human body and its spatiotemporal dynamics. Existing approaches often overlook these complexities, relying on general encoding methods that fail to distinguish different body parts and their dynamics, limiting precise semantic alignment. To address this, we propose WaMo, a novel wavelet-based multi-frequency feature extraction framework. It fully captures joint-specific and time-varying motion details at multiple resolutions for individual joint trajectories, extracting discriminative motion features to achieve fine-grained alignment with texts. WaMo has three key components: (1) Trajectory Wavelet Decomposition decomposes motion signals into frequency components that preserve both local kinematic details and global motion semantics. (2) Trajectory Wavelet Reconstruction uses learnable inverse wavelet transforms to reconstruct original joint trajectories from extracted features, ensuring the preservation of essential spatiotemporal information. (3) Disordered Motion Sequence Prediction reorders shuffled motion sequences to improve learning of inherent temporal coherence, enhancing motion-text alignment. Extensive experiments demonstrate WaMo's superiority, achieving 17.0\% and 18.2\% relative improvements in $Rsum$ on HumanML3D and KIT-ML datasets, respectively, outperforming existing state-of-the-art (SOTA) methods. Code is available at https://github.com/3DAgentWorld/WaMo/.

arXiv

Ultra-Short rPPG Estimation via Periodicity Guidance and Signal Reconstruction

Pei-Kai Huang, Ya-Ting Chan, Kuan-Wen Chen, Chiou-Ting Hsu, Xiaoding Wang, Md. Jalil Piran

Many remote Heart Rate (HR) measurement methods focus on estimating remote photoplethysmography (rPPG) signals from video clips lasting around 10 seconds but often overlook the need for HR estimation from ultra-short video clips. In this paper, we aim to accurately measure HR from ultra-short 2-second video clips by specifically addressing two key challenges. First, to overcome the limited number of heartbeat cycles in ultra-short video clips, we propose an effective periodicity-guided rPPG estimation method that enforces consistent periodicity between rPPG signals estimated from ultra-short clips and their much longer ground truth signals. Next, to mitigate estimation inaccuracies due to spectral leakage, we propose including a generator to reconstruct longer rPPG signals from ultra-short ones while preserving their periodic consistency to enable more accurate HR measurement. Extensive experiments on four rPPG estimation benchmark datasets demonstrate that our proposed method not only accurately measures HR from ultra-short video clips but also outperform previous rPPG estimation techniques to achieve state-of-the-art performance.

arXiv

AniCrafter: Customizing Realistic Human-Centric Animation via Avatar-Background Conditioning in Video Diffusion Models

Muyao Niu, Mingdeng Cao, Yifan Zhan, Qingtian Zhu, Weihang Ran, Yanhong Zeng, Xiao Sun, Zhihang Zhong, Yinqiang Zheng

Recent advances in video diffusion models have substantially enhanced character animation techniques. However, existing methods primarily depend on structural conditions, such as DWPose or SMPL-X, to animate character images, which limits their effectiveness in open-domain scenarios involving dynamic backgrounds or complex character-scene interactions. This study presents AniCrafter, a diffusion-based human-centric animation model designed to seamlessly integrate and animate a given character within open-domain dynamic backgrounds while adhering to specified human motion sequences. Built upon advanced Image-to-Video (I2V) diffusion architectures, the model introduces an innovative "avatar-background" conditioning mechanism that reformulates open-domain human-centric animation as a restoration problem, thereby achieving versatile, occlusion-aware animation results. Experimental evaluations demonstrate that the proposed approach outperforms current state-of-the-art methods and exhibits an exceptional capability in handling challenging scenarios. Codes and model are available at: https://github.com/MyNiuuu/AniCrafter

arXiv

So-Fake: Benchmarking and Explaining Social Media Image Forgery Detection

Zhenglin Huang, Xiangtai Li, Xi Yang, Bei Peng, Xiaowei Huang, Baoyuan Wu, Dacheng Tao, Ming-Hsuan Yang, Guangliang Cheng

Recent advances in AI-powered generative models have enabled the creation of increasingly realistic synthetic images, posing significant risks to information integrity and public trust on social media platforms. While robust detection frameworks and diverse, large-scale datasets are essential to mitigate these risks, existing academic efforts remain limited in scope: current datasets lack the diversity, scale, and realism required for social media contexts, while detection methods struggle with generalization to unseen generative technologies. To bridge this gap, we introduce So-Fake-Set, a comprehensive social media-oriented dataset with over 2 million high-quality images, diverse generative sources, and photorealistic imagery synthesized using 35 state-of-the-art generative models. To rigorously evaluate cross-domain robustness, we establish a novel and large-scale (100K) out-of-domain benchmark (So-Fake-OOD) featuring synthetic imagery from commercial models explicitly excluded from the training distribution, creating a realistic testbed for evaluating real-world performance. Leveraging these resources, we present So-Fake-R1, an advanced vision-language framework that employs reinforcement learning for highly accurate forgery detection, precise localization, and explainable inference through interpretable visual rationales. Extensive experiments show that So-Fake-R1 outperforms the second-best method, with a 1.3% gain in detection accuracy and a 4.5% increase in localization IoU. By integrating a scalable dataset, a challenging OOD benchmark, and an advanced detection framework, this work establishes a new foundation for social media-centric forgery detection research. The code, models, and datasets will be released publicly.

arXiv

Deepfake Media Generation and Detection in the Generative AI Era: A Survey and Outlook

Florinel-Alin Croitoru, Andrei-Iulian Hiji, Vlad Hondru, Nicolae Catalin Ristea, Paul Irofti, Marius Popescu, Cristian Rusu, Radu Tudor Ionescu, Fahad Shahbaz Khan, Mubarak Shah

We survey deepfake generation and detection techniques, covering all deepfake media types: image, video, audio and multimodal content. We identify various kinds of deepfakes and construct taxonomies of deepfake generation and detection methods, illustrating the important groups of methods. Next, we gather datasets used for deepfake detection and provide updated rankings of the best performing detectors on the most popular datasets. In addition, we develop a novel multimodal benchmark to evaluate deepfake detectors on out-of-distribution content. The results indicate that state-of-the-art detectors fail to generalize to deepfakes generated by unseen generators. Our project page and new benchmark are available at https://github.com/CroitoruAlin/biodeep.

arXiv

Contrastive Learning for Image Complexity Representation

Shipeng Liu, Liang Zhao, Dengfeng Chen, Zhanping Song

Quantifying and evaluating image complexity can be instrumental in enhancing the performance of various computer vision tasks. Supervised learning can effectively learn image complexity features from well-annotated datasets. However, creating such datasets requires expensive manual annotation costs. The models may learn human subjective biases from it. In this work, we introduce the MoCo v2 framework. We utilize contrastive learning to represent image complexity, named CLIC (Contrastive Learning for Image Complexity). We find that there are complexity differences between different local regions of an image, and propose Random Crop and Mix (RCM), which can produce positive samples consisting of multi-scale local crops. RCM can also expand the train set and increase data diversity without introducing additional data. We conduct extensive experiments with CLIC, comparing it with both unsupervised and supervised methods. The results demonstrate that the performance of CLIC is comparable to that of state-of-the-art supervised methods. In addition, we establish the pipelines that can apply CLIC to computer vision tasks to effectively improve their performance.

arXiv

CurveShift: Is Agent Progress Scalar? Separating Level from Shape

Hanwen Xing, Pengyun Wang, BingXu Meng, Kumail Alhamoud, Xiang Li, Jicheng Wang, Xin Yu, Xinyang Han, Xiaomin Li, Philip Torr, Yuexing Hao

Progress in large language models is often summarized using a single scalar measure, such as a time horizon, a latent ability estimate, or an aggregate benchmark score. These summaries capture the overall performance, but they do not test whether progress is distributed differently across task difficulty. We find that most of the apparent shift in gains toward harder tasks does not reflect a change in the shape of the difficulty-response curve. On METR time-horizon data, a single Rasch model with rising ability reproduces this pattern, so it is largely explained by ceiling effects rather than a qualitative change in capability. This echoes how the choice of metric can make claimed emergent abilities look like a property of the models themselves. We then identify a smaller hard-task effect that survives this control. Isolating it is difficult on agentic benchmarks, because newer models are usually run with newer agentic harnesses, so a gain on hard tasks cannot be assigned to the model or its scaffold. We break the confound with LiveCodeBench, a public competitive programming benchmark that runs no agentic scaffold while pairing dated models with an exogenous difficulty ordering. After accounting for the rise in overall ability, models released after September 2024 still gain on the hardest problems beyond what their easy and medium performance predicts, by about +0.40 logits under our most conservative assumption, raising the hard-problem solve rate from roughly 18% to 25%. The effect is led by the strongest reasoning models and holds for hard tasks that need only short reasoning, not autonomy over long horizons. We present this as a result specific to competitive programming, since our clean identification rests on a single coding benchmark. We release the LiveCodeBench Difficulty Panel (66 dated models x 1,055 problems) and our analysis code.

arXiv

Bayesian and Motivated Reasoning in AI Agents

Eddie Yang

AI agents increasingly perform open-ended tasks in settings where their conclusions can guide consequential decisions. We provide evidence that AI agents draw different conclusions from identical numerical data when the substantive framing changes. We demonstrate this behavior in high-stakes domains in medicine, election forensics, and geopolitical forecasting by holding the evidence fixed while changing the scenario in which the evidence appears. Across twelve agent-domain comparisons, agents' conclusions are strongly influenced by their prior beliefs. They are more likely to reach an affirmative conclusion when it is framed around a proposition they already regard as likely, while the reverse holds when the framing conflicts with their prior. The framing also changes how some agents work: they search more extensively, choose different analytical specifications, and evaluate the same evidence differently. These results identify a particular risk of delegating decision-making to AI agents, as their decisions may depend on prior beliefs that are neither specified in the task nor visible in the decision record.

arXiv

RMSWeb: Reflection, Failure-Mode Mining, and Salvage-DS for Web Agent Reinforcement Learning

Chengbo Liu, Lifang Zhou, Ruijie Yan, Pei Tan, Ao Sun, Haojun Huang, Guichun Hua, Sining Wei, Yining Chen, Yingying He, Yutao Xie

Compact web agents can reduce deployment cost, but training them poses challenges in both data collection and post-SFT reinforcement learning (RL). Successful trajectories are expensive to collect and often contain inefficient detours. After supervised fine-tuning (SFT), full trajectory corpora are dominated by routine states; moreover, when group-relative RL is applied to web actions, inadequately designed action-level rewards can yield weak or misleading relative updates, while groups rejected as unsuitable for such updates receive no fallback learning signal. We present RMSWeb, a three-part recipe for Qwen3-VL-Instruct at 8B and 32B. Reflection-conditioned retries increase collection yield and shorten successful trajectories; failure-mode mining concentrates offline RL on critical states exposed by the SFT policy; and Salvage-DS combines an action-semantic polarized reward, contrast-and-competence-gated dynamic sampling, and an action-only anchor for rejected groups. Policies trained with reflection-collected data use up to 19.7% fewer action steps on solved tasks. On WebVoyager, Online-Mind2Web, and WebTailBench, RMSWeb improves over SFT by 2.4-7.0 points at 8B and 1.2-7.7 points at 32B. Our 8B model also achieves the strongest reported Online-Mind2Web result among similarly sized open-weight models in our comparison and a leading reported accuracy-cost trade-off on WebVoyager and WebTailBench, with the caveat that external evaluation protocols differ.

arXiv

SeDeM: Selective Decompression of Hidden-State Memories for Long-Context Question Answering

Maryam Haghifam, Jason Cong, Yizhou Sun

Long-context inference with large language models (LLMs) is costly: self-attention during prefill scales quadratically with sequence length, and the key-value (KV) cache grows with the number of processed tokens. Larger context windows also do not ensure reliable evidence use. Context compression reduces this cost, but many soft-compression methods use LLMs as compressors and rely on compact memory tokens both to preserve information and to condition the decoder. We propose SeDeM, a selective decompression framework that decouples compact memory storage from decoder conditioning. An LLM extracts hidden states from a chosen intermediate Transformer layer, a lightweight compressor stores them as memory blocks, a query-conditioned selector selects relevant blocks, and a decompressor expands only the selected blocks into hidden states compatible with an intermediate decoder layer. Thus, the decoder avoids both full-context processing and direct generation from highly compressed memory slots. On four long-context QA benchmarks, SeDeM achieves higher QA scores than the evaluated compression baselines in both 1B and 3B same-backbone settings, and with the 3B backbone exceeds full-context fine-tuning on three datasets. The learned selector uses block-level evidence supervision during training. SeDeM also reduces online time-to-first-token and improves autoregressive decoding throughput relative to ICAE.

arXiv

Abstention as an Action Can Kill Both the Reward Gradient and the KL Anchor: Collapse Law and Repair for Error-Penalized Reinforcement Learning

Xujun Che, Yuchen Yuan, Weida Zhao, Chenyang Yu

Error-penalized scoring rules ($+1$ for a correct answer, $-λ$ for a wrong one, $0$ for abstaining) are increasingly prescribed against hallucination: a rational agent facing such a rule answers exactly when its correctness probability exceeds Chow's threshold $t^\ast=λ/(1+λ)$. We prove that a KL-anchored gradient learner can do the opposite. When abstention is a discrete action, the reward gradient and the anchor's restoring force are throttled by the same gate-saturation factor and die together: under explicit conditions (among them, blanket answering loses score in expectation and prompts share a bounded readout) the model drifts toward refusing everything, its mean training reward rising to zero like $1/t$ in training time $t$, so the curve reads as improvement while coverage collapses. The advantage estimator compounds the failure: in its sparse-answer regime, group normalization silently replaces every designed penalty with an effective penalty of one, moving the learned threshold from $t^\ast$ to $1/2$. The repair is structural: train a mandatory confidence report with a strictly proper score plus a correctness reward, and abstain only at deployment by thresholding the report. The always-emitted report has no gate to saturate, so no shared factor can kill its reward gradient and its anchor together, and its calibrated optimum is attracting. Simulations confirm every prediction, and experiments on language models at two scales confirm the mechanism live: the rule silences questions the models demonstrably still solve within ten optimizer steps, an ablation isolates the cause, and report-level training raises coverage, accuracy, and calibration together.

Keep browsing

More papers from top AI conferences

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