Search research

Test-time reinforcement learning adapts a model on its own unlabeled test set using majority-vote pseudo-labels and has shown strong results in mathematics. We show that this recipe collapses on medical multiple-choice QA: accuracy stagnates while output diversity rapidly declines. Through a controlled experiment that keeps the questions, model, and optimizer fixed while changing only the answer space, we trace this failure to answer-space structure rather than domain difficulty. In small answer spaces, incorrect rollouts often collide on the same wrong pseudo-label and reinforce it; in large answer spaces, they disperse and receive little reward. This diagnosis motivates PROSE, Process Reward Guided Self-Training, which rewards reasoning quality instead of answer agreement. PROSE scores each reasoning step with a medical process reward model, assigns the trajectory reward as the minimum score across steps, and enforces answer-format constraints. Without labels, PROSE substantially improves a general Llama model, surpassing purpose-built medical models and matching much larger systems. Because the process signal is internalized into the policy, the adapted model requires no reward model at inference and transfers its gains to unseen datasets. We further show that the minimum aggregation is essential: mean aggregation can be exploited, saturating the proxy reward while degrading accuracy.

Human decision-making is highly flexible -- some actions are taken immediately; others require longer deliberation. Language models have exhibited a similar capacity for adaptive "reasoning." However, transferring this capability to continuous control policies has been challenging, as directly reasoning in language space may lack the granularity for spatial understanding and precise motions. In this work, we show that reasoning for control policies can emerge by organizing information in an autoregressive latent space reminiscent of a memory palace, where retrieval is iterative and adaptive. Our method, Latent Memory Palace (LMP), formulates reasoning as variational inference with an autoregressive latent distribution. We derive a latent-space reinforcement learning technique to tractably optimize its variational lower bound. The resulting policy, LMP-$π$, achieves strong empirical performance in simulation and real-world domains while exhibiting interpretable, adaptive allocation of test-time compute. We further show that the same framework yields a variable-length action tokenizer, LMP-$\texttt{tok}$, which significantly improves the performance of downstream autoregressive policies. Together, these results present a new perspective on latent reasoning for control through the lens of variational inference.

Visual in-context learning has been proposed as a pathway towards dynamic models that can generate predictions based on a provided context and thereby can adapt to new vision tasks at test-time. Yet, the evaluation of the adaptation capabilities of these models has been limited to narrow setups that mainly mirror tasks or image domains from pre-training for which real adaptation is not required. We address this gap by constructing a broad Visual In-Context BEnchmark (VIBE) with a focus on diverse imaging domains and a wide range of tasks. With this, we are able to get a much clearer picture of the adaptive capabilities of visual in-context models when faced with new image- and task distributions. We stress test six models on $14$ datasets and $12$ tasks (in total, we explore $106$ dataset-task combinations) and compare them under a unified, reproducible evaluation protocol, in an one-shot setting. Our evaluation uncovers key insights on the state of visual in-context learning, including limitations, systematic failure modes and promising directions. To foster broader evaluation, we will openly release our VIBE toolkit.

Test-time training (TTT) adapts an LLM during generation by reading and updating request-owned state, such as fast weights, low-rank deltas, or streaming learner state. This breaks batched LLM serving, which assumes shared static weights: serial execution is correct but slow, while naive batching can corrupt request state. We formulate this problem as read-write TTT serving and present RW-TTT , which tags each decode step with its owner, version, and READ/WRITE effect, batches only compatible phases, and commits updates only to the owner. On one GPU with eight fast-weight InPlace-TTT streams, RW-TTT reaches 274.61 aggregate tok/s, 9.31x over sequential serving and 3.44x over per-stream replicas under the same memory budget. It preserves behavior on RULER, a long-context benchmark, and passes owner/version checks.

Existing deep learning models for Positron Emission Tomography (PET) image denoising often suffer from severe performance degradation under distribution shifts, fundamentally restricting their robust clinical deployment. This lack of generalization stems from the conventional paradigm of fixed-parameter models that cannot adapt to variations in test data (e.g., dose levels or scanner types) after training. To overcome this limitation and achieve robust generalization, we introduce U-TTT, a novel U-shaped model that integrates Test-Time Training (TTT) layers to dynamically adjust model parameters during inference through self-supervision, thereby adapting to the specific characteristics of each test instance. Furthermore, to comprehensively capture the complex degradations of 3D PET data, U-TTT features a dual-domain adaptation mechanism comprising a Spatial Test-Time Training (S-TTT) layer and a Frequency Test-Time Training (F-TTT) layer. The S-TTT layer captures and corrects spatial structural degradations, while the F-TTT layer suppresses global noise spectra and restores delicate high-frequency details. Extensive experiments demonstrate that U-TTT achieves state-of-the-art PET denoising performance and exhibits superior generalization under challenging distribution shifts, including both unseen dose levels and unseen scanners. Our code will be available at https://github.com/Yaziwel/U-TTT.

Medical time series (MedTS) signals such as electroencephalography (EEG) and electrocardiography (ECG) support many clinical applications. However, substantial subject-level heterogeneity often induces subject-level distribution shift, causing a fixed parameter set to generalize poorly to unseen individuals. Compared with domain adaptation methods that often depend on extra adaptation components or target-batch statistics, Test-Time Training (TTT) provides a more practical solution for sequential clinical data by enabling online adaptation from unlabeled test samples. However, many representative TTT methods require iterative inner-loop optimization, increasing test-time overhead. In this paper, we propose MedTS-TTT, a test-time training framework for medical time series modeling. MedTS-TTT is built upon Closed-Loop Self-Alignment Test-Time Training (CLSA-TTT) and a Gated Convolutional Backbone (GCB). CLSA-TTT constructs a token-level self-supervised target and performs a single-step fast-weight update for intra-layer closed-loop alignment, enabling rapid sample-wise adaptation without iterative inner-loop optimization. GCB combines CLSA-TTT-based fast adaptation and token-level fusion with a gated convolutional branch to balance local dynamic modeling and information-flow control. On 4 public datasets (2 EEG and 2 ECG) with subject-independent splits, MedTS-TTT achieves 11 top-1 rankings out of 12 evaluations across 9 baselines and 3 metrics. The code is publicly available at https://github.com/mingzhi-c/MedTS-TTT.

Optimization modeling is inherently hierarchical, requiring a precise sequence of symbolic commitments. Traditional learning-based automated optimization modeling methods improve modeling policies through large-scale annotated or curated training data, but are costly to adapt to new problem distributions. Meanwhile, one-shot generation remains brittle in hierarchical modeling, where early symbolic errors can propagate into invalid formulations. Test-time scaling offers a promising alternative by enabling structural exploration with additional instance-level computation; however, existing search-based methods typically rely on a fixed policy, causing repeated rollouts to inherit similar modeling biases and providing limited credit assignment for intermediate decisions. To address these limitations, we propose StarOR, a synergistic search-and-adaptation framework that couples MCTS with Test-Time Reinforcement Learning for optimization modeling. StarOR decomposes the modeling process into four stages and updates a transient LoRA adapter via GRPO at each non-terminal node. By using MCTS-generated siblings as local comparison sets, StarOR transforms search-time exploration into instance-specific policy refinement. Moreover, an unsupervised multi-faceted reward system provides fine-grained feedback for intermediate formulation decisions without ground-truth labels. Experiments across five optimization benchmarks show that StarOR achieves state-of-the-art performance even with a 4B backbone, outperforming existing methods and the frontier LLMs.

All-in-one adverse weather image restoration aims to remove multiple degradations, such as rain, haze, and snow, using a single unified model. Despite their broad applicability, existing methods typically compromise performance, delivering balanced but suboptimal results for individual degradation types. This issue becomes more pronounced when a domain gap exists between training and testing data. Motivated by the observation that modeling degradation patterns is more feasible than recovering clean content, we propose the Degradation Disentanglement and Transfer Network (DDTNet), which focuses specifically on degradation transfer. By disentangling degradation patterns from target-domain degraded images and transferring them to source domain clean images, DDTNet generates domain-adaptive paired training data. These pairs are then used to fine-tune restoration models, significantly enhancing their adaptability across diverse weather conditions and domains. The core of DDTNet is the Degradation Disentanglement Module (DDM), which comprises Degradation Coupled Attention (DCA) to capture both general and weather-specific features, thereby enabling effective disentanglement and transfer of degradation patterns. Experimental results demonstrate that DDTNet significantly and consistently improves existing all-in-one models across real-world deraining, desnowing, and dehazing datasets.

Generative personalization often suffers from the semantic collapsing problem (SCP), where a learned personalized concept overpowers the rest of the text prompt, causing the model to ignore important contextual details. To address this, we first analyze the underlying cause, revealing that the semantic drift responsible for SCP is not random but is concentrated within a specific low-dimensional subspace. We also discover that the personalization process perturbs the embedding of the original base concept, making it an unstable reference point. Based on these insights, we introduce Test-time Embedding Adjustment with Adaptive Subspace Projection (AdaptSP), a training-free method that uses the stable, pre-trained embedding as an anchor. AdaptSP isolates the semantic drift and projects it onto the identified subspace, performing a precise adjustment that mitigates SCP while maintaining the subject identity. Our experiments show that this targeted approach significantly improves prompt fidelity and contextual alignment.

Personalization in emotion recognition (ER) is essential for accurate interpretation of subtle and subject-specific expressive patterns. Recent advances in vision-language models (VLMs), such as CLIP, demonstrate strong potential for leveraging joint image-text representations in ER. However, existing CLIP-based methods either rely on CLIP's contrastive pretraining or use LLMs to generate descriptive text prompts, which can be noisy, computationally expensive, and often fail to capture fine-grained expressions, leading to degraded performance. In this work, Action Units (AUs) are leveraged as structured textual prompts within CLIP to model fine-grained facial expressions. AUs encode the subtle muscle activations underlying expressions, providing localized and interpretable semantic cues for more robust facial expression recognition (FER). We introduce CLIP-AU, a lightweight AU-guided temporal learning method that integrates interpretable AU semantics into CLIP. It learns generic, subject-agnostic representations by aligning AU prompts with facial dynamics, enabling fine-grained FER without CLIP fine-tuning or LLM-generated text supervision. Although CLIP-AU models fine-grained AU semantics, it does not adapt to subject-specific variability in subtle expressions. To address this limitation, we propose CLIP-AUTT, a video-based test-time personalization method that dynamically adapts AU prompts to videos from unseen subjects. By combining entropy-guided temporal window selection with prompt tuning, CLIP-AUTT enables subject-specific adaptation while preserving temporal consistency. Our experiments on three challenging video-based datasets, BioVid, StressID, and BAH, indicate that CLIP-AU and CLIP-AUTT outperform state-of-the-art CLIP-based FER and TTA methods.

Optimization-based latent reasoning improves large language model outputs by optimizing instance-specific continuous states at test time while keeping model parameters frozen. Existing methods, however, typically connect these states to the reasoning trajectory through decoded tokens, making sequence-level credit assignment indirect and obscuring how latent updates shape subsequent reasoning. We introduce GradCuit (gradient through circuit), which inserts optimizable latent states at a selected Transformer layer between the hidden representations of the prompt and the generated continuation. Causal self-attention provides every continuation-token log-probability with a differentiable path to every preceding latent state through the remaining Transformer blocks, enabling reward-weighted gradients from the entire continuation to be assigned directly to the latents. Across five instruction-tuned backbones, three reasoning benchmarks, and two answer formats, GradCuit achieves an average accuracy of 64.5%, outperforming chain-of-thought prompting by 6.6 percentage points and the strongest competing method by 2.4 points. GradCuit also demonstrates greater robustness: across seven learning-rate settings, it consistently outperforms LatentSeek while reducing the standard deviation of accuracy from 1.53 to 0.82, and even its random-walk variant remains competitive with LatentSeek. For interpretability, token-level gradient attribution reveals that latent influence concentrates on reasoning-connector tokens, while layer analysis identifies early-to-middle Transformer layers as the most effective optimization space. By directly optimizing internal reasoning from outcome feedback, GradCuit opens a new axis of robust and interpretable test-time scaling, where LLMs adapt how they reason rather than merely regenerate, sample, or rerank outputs.

Partial differential equations (PDEs) are foundational to modeling in science and engineering, but constructing reliable numerical solvers remains labor-intensive, demanding expert knowledge of discretization schemes, stability conditions, and boundary treatments. Recent work has begun to frame PDE solving as a code-generation task for large language models (LLMs), yet existing approaches operate primarily at inference time: relying on prompting, debugging, self-refinement, and test-time scaling rather than adapting the model itself. In parallel, reinforcement learning with verifiable rewards has emerged as a post-training paradigm for code and math reasoning, but its verifiers are typically binary: a compiler runs, or a test passes. Such signals discard the graded structure of scientific correctness, where two solvers may both execute and yet differ in solution accuracy by orders of magnitude. In this work, we introduce RLVP: Reinforcement Learning with Verifiable Physics, an RL post-training framework for multi-PDE solver code generation. RLVP addresses this verifiability gap with a hybrid verifier: hard program-validity checks ensure executability, while continuous physics rewards score function-space accuracy and PDE-residual consistency. A single policy is post-trained across diverse PDE families spanning hyperbolic, parabolic, elliptic, and incompressible-flow systems. RLVP improves over both pre-trained and supervised-only baselines on PDE benchmarks, and shows zero-shot improvement transfer to held-out PDEs. We show that a smaller LLM post-trained with RLVP can outperform prompting a frontier model on in-distribution PDE solver generation. The trained policy shows evidence of compositionality in numerical motifs: it recombines stencils, time-stepping schemes, and boundary-handling primitives learned from the PDEs used in training into generated solvers for unseen PDE problems.

Packet loss concealment (PLC) reconstructs audio packets that are missing at the receiver, usually with a trained model whose parameters remain fixed at deployment time. This treats the PLC model as static, even though each call or recording exposes signal-specific information through the packets that did arrive. We present TTT-PLC, a self-supervised test-time tuning framework that adapts existing PLC models using only those received packets. The method creates supervision by synthetically masking portions of the available signal, training the model to conceal them with its native PLC objective, and then using the adapted model to reconstruct the true packet losses. No clean reference signal, external adaptation data, or architectural modification is required. We study TTT-PLC in two deployment settings. In the non-causal setting, the received file is available before reconstruction, allowing repeated self-supervised adaptation passes and providing a per-file adaptation ceiling. In the causal setting, audio is streamed without revising emitted samples; adaptation is performed only on completed past blocks, and updated parameters affect only future audio. We instantiate the framework on two public PLC backbones, FRN, a recurrent full-band speech PLC model, and PARCnet, a hybrid autoregressive-neural model for networked music. Across these settings, the results show that pretrained PLC systems do not need to be treated as fixed at inference time, the still-observed portions of a lossy signal can provide an effective training signal for improving concealment on that same signal.

Pre-trained foundation models have demonstrated remarkable success in many domains, enabling a unified backbone to generalize across diverse downstream tasks. However, extending this paradigm to graph learning remains challenging due to the intrinsic mismatch between graph data and fixed architectural designs. In this work, we show that this limitation can be overcome via recurrent graph models. To achieve this, we conduct a systematic theoretical analysis, rigorously deriving step dependence as a necessary and sufficient condition for an adaptively convergent recurrent process. Building on this foundation, we propose AdaR, an Adaptive Recurrent graph model, empowering flexible test-time computing on various downstream tasks without changing model parameters. To enable adaptive inference, AdaR explicitly encodes normalized step information and representation-target relations into the recurrent updates. To ensure convergence of the recurrent process, AdaR employs gradient-based supervision signals that guide representation updates throughout the recurrence. Empirical results demonstrate that AdaR consistently outperforms strong baselines in both inductive and transductive settings.

Test-time alignment (TTA) aims to adapt models to specific rewards during inference. However, existing methods tend to either under-optimise or over-optimise (reward hack) the target reward function. We propose Null-Text Test-Time Alignment (Null-TTA), which aligns diffusion models by optimising the unconditional embedding in classifier-free guidance, rather than manipulating latent or noise variables. Due to the structured semantic nature of the text embedding space, this ensures alignment occurs on a semantically coherent manifold and prevents reward hacking (exploiting non-semantic noise patterns to improve the reward). Since the unconditional embedding in classifier-free guidance serves as the anchor for the model's generative distribution, Null-TTA directly steers model's generative distribution towards the target reward rather than just adjusting the samples, even without updating model parameters. Thanks to these desirable properties, we show that Null-TTA achieves state-of-the-art target test-time alignment while maintaining strong cross-reward generalisation. This establishes semantic-space optimisation as an effective and principled novel paradigm for TTA.

In real-world deployments of large language models (LLMs), balancing inference quality and computational cost has become a central challenge. Existing approaches tackle this trade-off along two largely independent dimensions: model routing, which switches among models of different scales to match request complexity, and test-time scaling (TTS), which adjusts inference-time compute within a fixed model for fine-grained control. However, this decoupled design introduces inherent limitations. Model routing yields coarse-grained, discrete performance changes due to the sparse set of model scales, while single-model TTS often encounters capacity ceilings and exhibits diminishing returns as compute increases. Moreover, treating the two mechanisms separately restricts adaptability in dynamic inference environments. To overcome these limitations, we introduce Unified Inference Scaling (UIS), which unifies model routing and TTS in a single optimization space. Building on this formulation, we propose UniScale, an online framework that models adaptive UIS as a contextual multi-armed bandit problem and learns inference policies via LinUCB. The framework incorporates efficiency-aware learning and cost modeling to ensure stable and scalable optimization over high-dimensional action spaces. Evaluation shows that UniScale effectively exploits the synergy in the UIS space to deliver a fine-grained and consistently better quality-cost trade-off across diverse, dynamic inference scenarios.

Biomedical imaging data presents enormous potential for deep learning models to predict invaluable properties, such as diseases and drug effects. However, unavoidable alterations of the technical conditions cause batch effects: variations between groups of samples that are not due to any biological signal of interest. Batch effects greatly hinder the generalization abilities of deep learning models, preventing their practical use in the real world. Unsupervised Domain Adaptation (UDA) methods have been proposed to mitigate batch effects, but they usually assume that the data is comprised of only one source domain and one target domain, whereas biological datasets are comprised of multiple domains, both at training and at inference time. While Batch Normalization-based test-time and meta-learning adaptation methods offer a promising mechanism for domain alignment, we show that existing approaches exhibit degraded performance under the usual inference scenarios of small target batch sizes and label shift. We address these limitations by leveraging negative control samples, which are consistently present in every experimental batch in biological datasets, as stable context for adaptation. We propose CS-ARM-BN, a meta-learning BN adaptation method that uses controls both during training and inference to stabilize domain statistics. We perform a suite of experiments of Mechanism-Of-Action (MoA) classification, a crucial task for drug discovery, on the large JUMP-CP imaging dataset. Our experiments show that CS-ARM-BN substantially improves robustness to batch size and class distribution shifts, enabling practical use of deep learning models for biomedical images.

Reconstruction of 3D objects from a single image is a challenging research problem in computer vision. The key challenge is the lack of critical information from viewpoints to complete 3D structures. Using an additional view may help to resolve the issue. However, there is no mechanism that can integrate the extra view into the single-view 3D reconstruction principle. We address this challenge by proposing ASV3D, a framework for adapting single-view 3D object reconstruction to test-time data with support from one additional image. We introduce two adaptation strategies: (i) a zero-shot adaptation scheme that leverages the auxiliary image to improve the reconstruction quality of an object without retraining, and (ii) an optimised adaptation scheme that further enhances visual fidelity and cross-view consistency via contrastive learning. We apply our ASV3D to improve two state-of-the-art single-view 3D reconstruction pipelines on both benchmark and real-world datasets. Results demonstrate that our approach consistently improves reconstruction accuracy and robustness under unconstrained multi-view inputs, outperforming the baselines in both quantitative metrics and human preference. We publish our code and the real-world object dataset in our project page at https://github.com/YNhuHuynh/ASV3D/tree/main.

Test-time reinforcement learning (TTRL) always adapts models at inference time via pseudo-labeling, leaving it vulnerable to spurious optimization signals from label noise. Through an empirical study, we observe that responses with medium consistency form an ambiguity region and constitute the primary source of reward noise. Crucially, we find that such spurious signals can be even amplified through group-relative advantage estimation. Motivated by these findings, we propose a unified framework, Debiased and Denoised test-time Reinforcement Learning (DDRL), to mitigate spurious signals. Concretely, DDRL first applies a frequency-based sampling strategy to exclude ambiguous samples while maintaining a balanced set of positive and negative examples. It then adopts a debiased advantage estimation with fixed advantages, removing the bias introduced by group-relative policy optimization. Finally, DDRL incorporates a consensus-based off-policy refinement stage, which leverages the rejection-sampled dataset to enable efficient and stable model updates. Experiments on three large language models across multiple mathematical reasoning benchmarks demonstrate that DDRL consistently outperforms existing TTRL baselines. The code will soon be released at https://github.com/yuyongcan/DDRL.

Test-time prompt tuning (TPT) enables adaptation on a single test instance, achieving improved accuracy but often sacrificing calibration performance. Most existing calibration methods introduce additional regularization terms to promote dispersion across text embeddings and reduce calibration error, yet these methods often suffer from a drop in accuracy. Motivated by the well-calibrated nature of zero-shot predictions, we propose CoTS, a simple yet effective post-hoc calibration method that preserves accuracy. Specifically, CoTS applies temperature scaling to minimize the confidence gap between adapted and zero-shot predictions. To fully exploit the potential of multiple augmentations during adaptation, we introduce a weak-strong ensemble strategy that further boosts accuracy. We then apply CoTS to this ensemble, termed E-CoTS, to maintain its well-calibrated property. Extensive experiments on diverse datasets and backbones show that our approaches effectively mitigate miscalibration without compromising primary accuracy. For instance, E-CoTS reduces the average expected calibration error of TPT from 11.90% to 5.38% on ImageNet variants, while even increasing accuracy from 60.74% to 62.95%. Moreover, when integrated with existing calibration methods, E-CoTS usually enhances both accuracy and calibration simultaneously.

Deep research agents augment large language models with external tools to answer complex, long-horizon questions through multi-turn reasoning. Learning from prior experience is crucial for continual improvement, yet existing methods either retrieve verbose task-specific traces that burden decision-making, or distill procedural skills that remain decoupled from downstream policy adaptation. We propose APEx, a hierarchical experience utilization framework that organizes interaction history into instance-level trajectory memories and category-level procedural skills, and couples them through a closed-loop architecture of Executor, Distiller, and Planner. The three modules are optimized via a three-stage alternating GRPO training paradigm, enabling reward-guided skill distillation rather than fixed-prompt generation. At test time, distilled skills serve as procedural priors for online Planner adaptation through skill-guided test-time reinforcement learning, allowing ground-truth-free self-improvement with skill-alignment regularization to prevent policy drift. Experiments on 7 benchmarks demonstrate that APEx achieves state-of-the-art performance, surpassing GPT-5.4 by 14.7 points and the strongest memory-augmented baseline by 3.0 points.

Mobile GUI agents can operate apps through pixel perception and touch actions, making them a promising interface for collecting and improving long-horizon mobile interaction policies. However, real trajectories are difficult to obtain for sensitive apps and privacy-critical operations. At the same time, existing simulated environments are costly to scale up, and GUI world models still suffer from unstable generation, limited modality coverage, and inconsistent action-transition logic. To address these limitations, we propose AppDeltaWorld, a transition-grounded delta code world model that predicts the next GUI as a reachable code update rather than as an unconstrained image or text description. AppDeltaWorld retrieves app-specific Level-1 HTML references under an action-transition constraint, generates Level-2 executable HTML conditioned on the current screen, action, predicted next-screen text, and retrieved structure, and inserts generated visual assets into image slots before browser rendering. As a world model, AppDeltaWorld achieves the highest fidelity on CMGUIBench-500 under Code2World evaluation, with clear gains in structural layout and UI element reconstruction over image-only and code-only baselines. As a training environment, AppDeltaWorld supports filtered closed-loop SFT data construction that, when combined with public supervision, enables AppDeltaAgent to achieve state-of-the-art performance on AndroidLens and consistent gains on MobileGym and MobileWorld. Moreover, world-model-based test-time reinforcement learning enables policy adaptation and shows further improvements without additional interaction with real apps.

Conformal predictions have attracted significant attention in the field of uncertainty quantification, mainly because of their strong marginal coverage guarantees. Full conditional guarantee is not an attainable goal, a well known fact in conformal predictions literature. As a result, several approaches have tried to approximate this behavior by adapting the conformal sets of test-time samples according to their similarity to calibration examples. Although the latter has gained traction and shown impressive performances for regression problems, its application to image classification remains under-explored. We conduct an extensive benchmarking on natural image classification tasks with vision-language models (VLMs), using our open source implementation of a recent localized conformal prediction algorithm. We show that straightforward usage of the cosine similarity between test-time and calibration visual features, an intuitive choice for VLMs, is not sufficient to improve over the non-local baselines. In response, we propose a simple non-linear transformation of the cosine similarities, which conserves marginal coverage guarantees and achieves statistically significant mean set sizes reduction. Code is available at https://github.com/cfuchs2023/lcp-vlm/.

Preference modeling plays a central role in reinforcement learning from human feedback (RLHF), enabling large language models (LLMs) to align with human values. However, most existing approaches assume a universal reward function, neglecting the diversity and heterogeneity of human preferences. To address this limitation without additional annotation costs, recent work has proposed learning multiple preference components from binary data and combining them to model individual preferences. Nevertheless, these components often fail to capture coherent and disentangled patterns, limiting their interpretability and effectiveness for personalization. In this work, we propose a sparse Mixture-of-Experts (MoE) reward model that encourages sparse routing and expert diversity during training on binary preference data. Across controlled and real-world experiments, sparse MoE learns interpretable routing patterns and specialized experts. It also improves test-time personalization, and post-adaptation shifts in expert weights provide a qualitative lens for analyzing how the model adapts to personalized preferences.

Wireless foundation models are rapidly emerging as a key enabler of AI-native communication systems, yet a fundamental question remains unanswered: how large should these models be? We present a principled, physics-grounded answer, showing that the intrinsic dimensionality (dNL, the nonlinear manifold dimension of the channel) acts as the fundamental bottleneck, defining the scaling ceiling once a data-sufficient regime is reached. This dimensionality is not a design choice but a physical constraint: Maxwell's equations, finite scatterers, and antenna aperture inherently constrain wireless propagation environments to a limited number of degrees of freedom -- spanning 5-35 across both real-world OTA measurements and 3GPP-standardized channel models we evaluate -- orders of magnitude below the ~1,000-dimensional semantic space of language. As a consequence, we propose a scaling framework for wireless AI: taking NTN satellite channels as a representative case (dNL ~= 14), scaling gains diminish rapidly beyond ~30 million parameters, entering a stochastic asymptote above 70M where a further 1.6x increase (96M->150M) yields only 0.52 dB. Beyond this ceiling, inference-time adaptation via pilot-aided test-time training (TTT) is far more effective: a compact 12M-parameter model surpasses a static 96M model by 9.9 dB (NMSE, SNR = 20 dB) / 7.6 dB (MCM, SNR = 10 dB) at one-eighth the parameters. With dNL distributions validated across real-world indoor massive MIMO measurements, our scaling laws and TTT gains are demonstrated through NTN satellite simulations, reframing wireless AI design: channel geometry -- not model size -- fundamentally governs the scaling laws of physical-layer wireless AI.

Visual Foresight VLA (VF-VLA) has become a prominent architectural choice in the recent VLA due to its impressive performance. Nevertheless, the inherent design of VF-VLA makes it particularly vulnerable to out-of-distribution (OOD) shifts. Because the quality of action directly depends on the accuracy of the predicted future visual information, OOD conditions affect both stages at once. To address this vulnerability, we propose Test-Time Training Visual Foresight VLA ($T^3$VF), a test-time training approach motivated by the observation that the predicted future image and its subsequent observation form a natural supervision pair. To further address the practical challenges that arise from indiscriminate test-time updates, we introduce an adaptive update filtering mechanism. Empirically, $T^3$VF mitigates the OOD vulnerability of VF-VLA at a modest additional inference cost, without requiring any architectural modification or auxiliary modules.

Large language models (LLMs) have shown increasing promise in generating functionally correct register-transfer-level (RTL) hardware designs. Recent systems improve further through EDA-integrated reinforcement learning with syntax, simulation, and PPA rewards, but train a general RTL generator before deployment while test-time approaches search with a frozen policy. We instead perform reinforcement learning at test time, allowing the LLM policy to adapt to executable EDA feedback for the specific RTL problem at hand. We propose TTT-RTL, to our knowledge the first per-design test-time training framework that closes the loop between an LLM policy and an EDA pipeline for RTL optimization. TTT-RTL samples candidate implementations, verifies them through syntax checking and simulation, scores valid designs using synthesis-derived PPA product, reuses high-reward variants through a PUCT-indexed design-state pool, and updates the policy with an entropic policy-gradient objective. To stabilize policy updates under sparse or plateaued rewards, we introduce an adaptive KL-budget controller that adjusts the entropy constraint using reference KL, effective sample size, and reward saturation signals. On RTLLM v2.0 under Nangate 45nm, TTT-RTL reduces the geometric-mean PPA product by 65.1% over the reference, outperforming the strongest published frozen-policy agent baseline at 26.1%. On an industrial XuanTie C910 FPU leading-zero-anticipation unit under Sky130, TTT-RTL achieves a 59.4% ADP reduction, and ablations confirm that policy adaptation, state reuse, and KL-budget control each contribute. These results suggest that test-time training with executable EDA feedback can move LLM-based RTL generation beyond functional correctness toward physically optimized hardware.

By incorporating test-time compute scaling, large reasoning models (LRMs) can solve complex problems through explicit chain-of-thought (CoT) reasoning processes. However, they often suffer from overthinking, resulting in redundant token outputs and degraded accuracy. Current methods to mitigate this issue remain limited: training-based approaches require substantial computational resources, while training-free methods rely on well-crafted prompts or unreliable confidence signals. In this work, we investigate early stopping from the perspective of attention distributions and propose a simple method, ASAG, which infers the model's reasoning state and adaptively adjusts the generation strategy. The proposed framework is training-free and plug-and-play, enabling seamless integration into existing LRMs. Extensive experiments on nine benchmarks demonstrate consistent improvements across mainstream LRMs with varying parameter scales, including the DeepSeek-R1-Distill and Qwen3 series. Specifically, ASAG improves average accuracy by 3.2% while reducing the number of generated tokens by nearly 40% across all reasoning tasks on Qwen3-8B.

Visual in-Context Learning (VICL) aims at making progress towards adaptive vision models, that can -- based on a few examples -- adapt to a new task at test-time. With the history of in-context learning in natural language processing research, where large, parameter-heavy models are in use, one pathway that current VICL methods take is model- and data-scaling as key ingredients. Yet, it is not clear, whether these ingredients are the key for in-context learning to take shape in vision models. To stress-test such large models, we challenge them with an extreme counterexample: we train a tiny visual in-context model with merely $1$ million parameters and a modest amount of $70,000$ images. We compare the results of this severely capacity capped tiny model to $7,000\times$ larger VICL models in different adaptive settings, (1) on image data with small distribution shifts, (2) on unseen task encodings and (3) on a completely new task, i.e., the setting VICL envisions. With the chasm of training resources between the tiny- and large models, our experiments showcase a lack in how adaptive capabilities are measured, with respect to how tasks are encoded, which tasks were used in pre-training and the choice of metrics. These gaps in current VICL benchmarking underscore a need for innovation in evaluation of adaptive capabilities.

Ensuring the security of reinforcement learning (RL) models is critical, particularly when they are trained by third parties and deployed in real-world systems. Attackers can implant backdoors into these models, causing them to behave normally under typical conditions, but execute malicious behaviors when specific triggers are activated. In this work, we propose Plan2Cleanse, a test-time detection and mitigation framework that adapts Monte Carlo Tree Search to efficiently identify and neutralize RL backdoor attacks without requiring model retraining. Our approach recasts backdoor detection as a planning problem, enabling systematic exploration of temporally extended trigger sequences while maintaining black-box access to the target policy. By leveraging the detection results, Plan2Cleanse can further achieve efficient mitigation through tree-search preventive replanning. We evaluated our method in competitive MuJoCo environments, simulated O-RAN wireless networks, and Atari games. Plan2Cleanse achieves substantial improvements, increasing trigger detection success rates by more than 61.4 percentage points in stealthy O-RAN scenarios and improving win rates from 35% to 53% in competitive Humanoid environments. These results demonstrate the effectiveness of our test-time defense approach and highlight the importance of proactive defenses against backdoor threats in RL deployments. Our implementation is publicly available at https://github.com/rl-bandits-lab/RL-Backdoor.