University of Oxford（牛津大学）

Sparse autoencoders (SAEs) decompose language model activations into sparse features, but standard SAEs encode each token independently and do not expose information that persists across a sequence. We introduce Persistent Sparse Autoencoders (Persistent SAEs), which extend standard SAEs by learning a persistence coefficient for each feature, allowing the model to learn which features should persist and for how long. Our experiments show that they retain competitive reconstruction quality while learning a spectrum of feature timescales: fast features behave as locally interpretable detectors, whereas slow features concentrate topic-level information in a persistent state. Moreover, as shown in a prompt-injection monitoring case study, slow features preserve detection signals and remain causally effective over long contexts. These results suggest that Persistent SAEs open up new opportunities for interpreting and monitoring language models through persistent semantic representations.

Autoregressive multimodal large language models (MLLMs) suffer from error snowballing: a single incorrect inference early in a chainof-thought (CoT) trace corrupts all downstream reasoning. We find that in state-of-the-art open-source MLLMs, once the first error occurs, the reasoning cascades into failure across all remaining steps in 65% of such cases (a metric we term the snowball rate). Existing mitigations-sampling multiple chains, post-hoc self-verification, or full program synthesis-either lack symbolic grounding, catch errors too late, or sacrifice the flexibility of natural language reasoning. We propose Constraint-Anchored Reasoning Traces (CART), a neuro-symbolic framework that trains MLLMs to interleave natural language reasoning steps with symbolic constraint assertions: lightweight, machine-checkable statements about visual content (e.g., count(red_objects) = 3). A dual-pronged Constraint Propagation Module-combining a learned neural grounding head with Boolean Constraint Propagation-continuously verifies these anchors against extracted visual features and checks their mutual logical consistency. When a contradiction is detected, a backtrack controller halts generation and reverts to the last consistent checkpoint, preventing error propagation. A variable-frequency emission mechanism allows the model to adaptively control anchor density, avoiding trace bloat. We construct 218K training instances by augmenting GQA, CLEVR-CoGenT, and VCR with ground-truth constraint annotations derived from scene graphs, and fine-tune open-source MLLMs (LLaVA-NeXT, Qwen2-VL) via LoRA. On five benchmarks, CART reduces the snowball rate from 0.65 to 0.14, improves GQA accuracy by +4.6 percentage points over trainingonly baselines, and achieves 89.1 F1 on POPE-all with at most 18% inference overhead.

This paper investigates whether multimodal LLMs can evaluate local brushstroke quality in calligraphy and generate educationally useful natural language feedback. We construct an evaluation framework in which three multimodal LLMs (GPT-4o, Claude Sonnet 4, and Gemini 2.5 Flash) assess before-after image pairs of calligraphic works using a five-point ordinal scale, and compare their outputs against scores assigned by three expert calligraphers. We additionally examine a Retrieval-Augmented Generation (RAG) variant of Claude as a preliminary condition. Results show that all models achieve useful levels of absolute score accuracy (MAE), with GPT-4o performing best (MAE = 0.885). However, none of the models produce statistically significant overall rank correlations with human experts (Kendall's tau). Vocabulary analysis of generated rationales reveals characteristic evaluative biases in each model, and RAG is shown to improve rank correlation while worsening absolute accuracy, constituting an important negative result for text-based rule injection.

AI systems are increasingly used to support forward-looking scientific judgment, but it remains unclear whether they can form reliable expectations about future scientific advances. Here we show that strong scientific reasoning does not reliably translate into accurate forecasting of future scientific advances. To study this question, we introduce CUSP, a temporally grounded evaluation suite for event-level scientific forecasting across eight scientific disciplines. Across six frontier AI models, we observe a striking asymmetry in forecasting performance together with systematic error patterns. Models often identify plausible mechanisms underlying future scientific advances, yet perform near chance on feasibility assessment, generate solution strategies that only weakly align with realized advances, and systematically predict scientific advances later than they become publicly observable. Providing additional pre-cutoff scientific knowledge improves performance but does not eliminate these forecasting limitations. These findings suggest that current AI systems possess substantial retrospective scientific competence but limited forward-looking predictive capability. Scientific forecasting should therefore be evaluated as a complementary dimension of AI scientific capability when deploying AI systems for research prioritization and scientific decision-making.

Bayes-filtered transformers are transformers meta-learned on sequences from a prior predictive distribution to approximate the corresponding posterior predictive distribution. They output total predictive uncertainty in a single forward pass but never explicitly represent a posterior distribution, making the standard route to separating aleatoric from epistemic uncertainty unavailable. We address this challenge through the lens of Bayesian predictive inference (BPI). Our main result is a predictive Central Limit Theorem (CLT) for supervised settings under conditions that are among the weakest known in the BPI literature. The CLT characterises the posterior of the limiting predictive distribution given an observed context as asymptotically Gaussian; the variance of this Gaussian quantifies epistemic uncertainty. We apply the framework to TabPFN, a Bayes-filtered transformer that is a state-of-the-art foundation model for tabular prediction. The resulting credible bands achieve near-nominal frequentist coverage as context length grows, and the decomposition largely matches standard desiderata: epistemic uncertainty shrinks with context length and is highest in sparsely observed regions within the span of the context data, while aleatoric uncertainty dominates near decision boundaries where classes overlap.

Cement production underpins global infrastructure but contributes approximately 7% of anthropogenic CO2 emissions, making accurate monitoring of production facilities essential for sustainable development. Existing remote sensing approaches rely predominantly on thermal signatures from kiln operations, which can be confounded by background heat sources and fail to capture persistent environmental alterations. This study introduces a physics-informed methodology that exploits multi-spectral features, particularly a geological Short Wave Infrared (SWIR) ratio, to detect soil property changes induced by sustained industrial heat emissions. This work proposes FusionNet, an intermediate multi-spectral data fusion framework that integrates Thermal Infrared (TIR) and SWIR inputs through a specialised backbone, embedding differential signal processing priors within a convolutional layer, mixed pooling, and wider receptive field. Systematic ablation studies confirm that each architectural component contributes to performance gains, with DGCNN achieving a 4.1-6.8% accuracy improvement over conventional CNNs. On the SWIR ratio dataset, FusionNet attains a maximum of 90.6%, outperforming state-of-the-art baselines across five spectral configurations and exceeding the strongest unimodal model by 1.1%. Transfer learning experiments reveal that ImageNet pretraining degrades TIR and SWIR performance, underscoring the importance of modality-aware training for cross-spectral applications. Overall, the results demonstrate that combining physics-aware feature selection with principled deep learning architectures enables robust, high-accuracy detection of cement production facilities, offering a reliable framework for industrial infrastructure monitoring

Soft Q-learning has emerged as a versatile model-free method for entropy-regularised reinforcement learning, optimising for returns augmented with a penalty on the divergence from a reference policy. Despite its success, the multi-step extensions of soft Q-learning remain relatively unexplored and limited to on-policy action sampling under the Boltzmann policy. In this brief research note, we first present a formal $n$-step formulation for soft Q-learning and then extend this framework to the fully off-policy case by introducing a novel Soft Tree Backup operator. Finally, we unify these developments into Soft $Q(λ)$, an elegant online, off-policy, eligibility trace framework that allows for efficient credit assignment under arbitrary behaviour policies. Our derivations propose a model-free method for learning entropy-regularised value functions that can be utilised in future empirical experiments.

Fine-tuned LLMs can covertly encode prompt secrets into outputs via steganographic channels. Prior work demonstrated this threat but relied on trivially recoverable encodings. We formalize payload recoverability via classifier accuracy and show previous schemes achieve 100\% recoverability. In response, we introduce low-recoverability steganography, replacing arbitrary mappings with embedding-space-derived ones. For Llama-8B (LoRA) and Ministral-8B (LoRA) trained on TrojanStego prompts, exact secret recovery rises from 17$\rightarrow$30\% (+78\%) and 24$\rightarrow$43\% (+80\%) respectively, while on Llama-70B (LoRA) trained on Wiki prompts, it climbs from 9$\rightarrow$19\% (+123\%), all while reducing payload recoverability. We then discuss detection. We argue that detecting fine-tuning-based steganographic attacks requires approaches beyond traditional steganalysis. Standard approaches measure distributional shift, which is an expected side-effect of fine-tuning. Instead, we propose a mechanistic interpretability approach: linear probes trained on later-layer activations detect the secret with up to 33\% higher accuracy in fine-tuned models compared to base models, even for low-recoverability schemes. This suggests that malicious fine-tuning leaves actionable internal signatures amenable to interpretability-based defenses.

Myocardial infarction (MI) remains a leading cause of mortality worldwide. Echocardiography (Echo) is a widely available modality for MI assessment, where regional wall motion abnormality is a key indicator. Prior learning based methods for myocardial motion analysis often use handcrafted descriptors or densely supervised estimation, but the need for extensive annotation limits applicability. Foundation models have recently improved vision-based Echo analysis; however, most methods operate on single views and segment-level localization remains unreliable under view-dependent ambiguity, especially in apical views. To address this, we propose MCF-Net, a novel motion-guided multi-view fusion framework that fuses myocardial motion cues with foundation model representations to localize infarction. Visual features are extracted using EchoPrime, a pretrained Echo foundation model shared across dual views. Cardiac motion is modeled with extremely sparse supervision: a single annotated template frame is transferred across videos to initialize point tracking, avoiding dense labels. Motion-derived segment-aware soft masks provide coarse spatial priors that selectively enhance features for challenging myocardial segments. A motion-conditioned fusion mechanism then integrates motion and vision across views, refining predictions without overriding strong appearance cues. On segment-level MI localization, MCF-Net achieves 72.4\% F1 and 84.9\% accuracy, outperforming state-of-the-art motion-only, vision-only, and fusion baselines.

Recent advances in Vision-Language Models (VLMs) have significantly improved image geo-localization, yet existing models remain susceptible to landmark bias, causing them to overlook geographical cues or form spurious correlations, ultimately resulting in inaccurate localization. To systematically investigate this issue, we first design two quantitative metrics, Bias Intensity (BI) and Bias Harmfulness (BH), to characterize the impact of landmarks exerted on model reasoning, and establish a comprehensive benchmark, LandmarkBias-3K. To mitigate landmark bias, we further propose an evidence-driven reasoning framework, HoloGeo, to improve the reliability of geo-localization. HoloGeo is supported by a high-quality dataset, BF-30k, annotated with structured multi-evidence bias-free reasoning chains. By incorporating multi-dimensional rewards, HoloGeo explicitly encourages balanced attention over diverse visual cues and achieves evidence-driven joint reasoning. Extensive experiments demonstrate that HoloGeo not only maintains excellent performance on IM2GPS3K and YFCC4k but also significantly outperforms existing open-source VLMs on LandmarkBias-3K, validating its effectiveness for robust geospatial reasoning.

Direct hand-driven teleoperation maps an operator's hand motion to robot end-effector commands at every frame, enabling precise control, but it requires constant monitoring and correction during approach, grasp, and placement, which can be slow and fatiguing. For repetitive pick-and-place tasks, supervisory (goal-based) teleoperation simplifies this process: the operator specifies goals/waypoints, and the robot executes the motion using planning algorithms. Yet, this introduces latency, as the robot must wait for the next command before it can plan and act. "How can we reduce robot reaction time while lowering operator workload?" To tackle this question, we present AHEAD, a real-time VR teleoperation system that anticipates operator intent to enable proactive, hand-driven control. In a digital twin, the operator performs pick-and-place naturally, using hand motion to convey high-level commands rather than a continuous robot trajectory. AHEAD processes a short window of 3D hand and head signals together with scene context through an attention-based classifier to predict the intended grasp object and placement slot. A state machine converts intent predictions into stable robot goals, enabling early motion while remaining stable under noisy predictions and corrective hand movements. AHEAD's intent prediction module achieves Top1 accuracy: 76% for grasp objects and 76% for target slots. Moreover, our user study shows AHEAD reduces robot reaction latency by 0.6 s (object) and 1.4 s (slot) relative to baselines. Participants also reported lower operator load, indicating faster robot responses while maintaining low operator effort in practice.

Recent work has raised concerns about the influence of sycophantic AI on user judgment and relationships. One proposed mitigation, which has received regulatory attention, is to warn users about potentially harmful AI behaviors such as sycophancy. In a preregistered experiment in which participants (N = 2,610) discussed real interpersonal conflicts with an AI system, we test whether warning labels mitigate sycophancy's influence. We find that a basic AI disclosure (``This chatbot is AI'') has no detectable effect. Labeling the system as sycophantic (``...may agree with you and validate you even when you are wrong...'') does shift users' perceptions, reducing perceived objectivity and trust, but it does not reliably reduce sycophancy's influence on users' self-perceived rightness or their willingness to repair the conflict. Our results reveal a gap between AI perception and AI influence: by shifting perception without reducing influence, warning-based interventions may offer a false sense of protection. Addressing the harms of sycophancy will therefore require understanding the specific mechanisms through which it shapes judgment, and improving model behavior itself.

Research metrics use attention as evidence of societal impact. Yet attention serves as evidence only once interpreted, and its meaning depends on its contextual structure, not on volume alone. Altmetrics represents signals in isolation, keeping a count of the attention an output received, or a sequence of when. We address this with attention flows, representations that situate an output's attention in the contexts through which it is distributed. To evaluate the flow, we build a benchmark of analogy queries, each testing whether the relationship between two outputs, applied to a third, yields a fourth. The count and sequence baselines fail to recover these relationships, whereas flows learned as dynamic contextualised representations recover them. The recovered structure also survives partial observation and rests on its contexts instead of volume. These findings support attention represented as contextually structured for research evaluation.

Large language models (LLMs) are increasingly used to answer health-related questions and support healthcare workflows, yet evidence for their safety still relies heavily on static benchmarks that can rapidly become obsolete or be optimized against. Here we introduce a Dynamic, Automatic, and Systematic (DAS) red-teaming audit framework that continuously stress-tests LLMs for health across four safety-critical axes: robustness, privacy, bias/fairness, and hallucination/factual inaccuracies. Validated against board-certified clinicians with high concordance, a suite of adversarial agents autonomously mutates health-related test cases to uncover vulnerabilities in real time. Applying DAS to 15 proprietary and open-source LLMs revealed a profound gap between high static benchmark performance and low dynamic reliability--the "Benchmarking Gap". Despite median MedQA accuracy exceeding 80\%, 94\% of previously correct answers failed under dynamic robustness testing. This brittleness generalized to the realistic, open-ended HealthBench dataset, where top-tier models exhibited failure rates exceeding 70\% and sharp shifts in model rankings across evaluations, suggesting that high scores on established static benchmarks may reflect superficial memorization. We observed similarly high failure rates across other domains: privacy leaks were elicited in 86\% of scenarios, cognitive-bias priming altered recommendations in 81\% of fairness tests, and hallucination rates exceeded 74\% in widely used models. By converting LLM safety evaluation for health from a static checklist into a living adversarial audit, DAS provides a scalable framework for surfacing latent risks before such systems are deployed in consumer-facing health assistants, clinician-facing tools, and broader healthcare workflows. Code is available at https://github.com/JZPeterPan/DAS-Medical-Red-Teaming-Agents.

Monte Carlo Tree Search is a cornerstone algorithm for online planning, and its root-parallel variant is widely used when wall clock time is limited but best performance is desired. In environments with continuous action spaces, how to best aggregate statistics from different threads is an important yet underexplored question. In this work, we introduce a method that uses Gaussian Process Regression to obtain value estimates for promising actions that were not trialed in the environment. We perform a systematic evaluation across 6 different domains, demonstrating that our approach outperforms existing aggregation strategies while requiring a modest increase in inference time.

We propose the AIMO Interpretability Challenge, a competition on distinguishing robust from spurious reasoning in frontier mathematical language models based on the models' internal mechanisms. The challenge is motivated by a central limitation of standard reasoning benchmarks: strong final-answer accuracy does not reveal whether a model relies on stable reasoning mechanisms or exploits brittle reasoning shortcuts. Building on AI Mathematical Olympiad (AIMO) problems and submissions, together with resources from the Fields Model Initiative, the competition will provide (1) newly-published olympiad-level math reasoning problems and their symbolic representations, allowing generation of novel functional variants, (2) access to frontier reasoning models, and (3) assessments of models' adversarial robustness on these problems. Participants will use these resources, along with our computing infrastructure support, to develop methods for identifying which models solve problems robustly. Our competition will also create a new, open robustness benchmark and baseline systems, aiming to provide a lasting foundation for standard benchmarking in mathematical reasoning and interpretability. Scientifically, the competition connects interpretability and generalization research around a central question in AI research: can we determine if, and to what extent, the decision-making of frontier AI models is generalizable and thus, reliable?

Physics-informed learning promises data-efficient and stable dynamics prediction, yet its strongest geometric guarantees have largely remained confined to closed conservative systems. This excludes robotic systems of interest, where actuation, dissipation, and constraints exchange energy and momentum with the environment. We introduce CaLiSym, a lightweight framework that extends symplectic learning to such systems by changing where the geometric prior is imposed. Rather than enforcing symplecticity on the measured state, CaLiSym embeds the state and its ports into a lifted phase space, where the dynamics evolve through a symplectic map. The lift is explicit and algebraic, requiring neither recurrent latent states, transformer decoders, implicit optimization, nor inference-time numerical integration. We instantiate the framework with SympNet predictors and introduce GRB-SympNet, a B-spline variant combining approximation with exact symplectic structure. Experiments on a controlled dissipative double pendulum, a real-world quadrotor, and a contact-constrained real-world quadruped demonstrate the lowest out-of-distribution autoregressive rollout error across systems, improving by up to 69.5% while using fewer parameters and up to 85x fewer floating-point operations per step than sequence-model baselines. The lifted dynamics preserve the symplectic form to numerical precision, extending symplectic learning beyond conservative mechanics toward real-world robotics.

Chain-of-thought prompting has popularized step-by-step reasoning in large language models, yet model performance still degrades as problem complexity and context length grow. By decomposing difficult tasks with long contexts into shorter, manageable ones, recent multi-agent paradigms offer a promising near-term solution to this problem. However, the fundamental capacities of such systems are poorly understood. In this work, we propose a theoretical framework to analyze the expressivity of multi-agent systems. We apply our framework to three algorithmic families: state tracking, recall, and $k$-hop reasoning. We derive bounds on (i) the number of agents required to solve the task exactly, (ii) the quantity and structure of inter-agent communication, and (iii) the achievable speedups as problem size and context scale. Our results identify regimes where communication is provably beneficial, delineate tradeoffs between agent count and bandwidth, and expose intrinsic limitations when either resource is constrained. We complement our theoretical analysis with a set of experiments on pretrained LLMs using controlled synthetic benchmarks. Empirical outcomes confirm the tradeoffs between key quantities predicted by our theory. Collectively, our analysis offers principled guidance for designing scalable multi-agent reasoning systems.

Spatio-temporal fetal brain atlases are important for characterizing normative neurodevelopment and identifying congenital anomalies. However, existing atlas construction pipelines necessitate days for slice-to-volume reconstruction (SVR) to generate high-resolution 3D brain volumes and several additional days for iterative volume registration, thereby rendering atlas construction from large-scale cohorts prohibitively impractical. We address these limitations with INFANiTE, an Implicit Neural Representation (INR) framework for high-resolution Fetal brain spatio-temporal Atlas learNing from clinical Thick-slicE MRI scans, bypassing both the costly SVR and the iterative non-rigid registration steps entirely, thereby substantially accelerating atlas construction. Extensive experiments demonstrate that INFANiTE outperforms existing baselines in subject consistency, reference fidelity, intrinsic quality and biological plausibility, even under challenging sparse-data settings. Additionally, INFANiTE reduces the end-to-end processing time (i.e., from raw scans to the final atlas) from days to hours compared to the traditional 3D volume-based pipeline (e.g., SyGN), facilitating large-scale population-level fetal brain analysis. Code: https://github.com/hu2274898/INFANiTE

Diffusion models have proven to be a flexible and effective framework for modelling probability distributions on finite-dimensional spaces. However, many physical modelling problems such as time series are naturally described over function spaces. In this work we apply diffusion models to such stochastic processes. To do so we consider a spectral representation of the data, obtained using a kernel, thereby dissociating the stochastic part of the processes from their space-time structure. As a result, the stochasticity of the processes is entirely encoded in the spectral coefficients, which we truncate and model using standard finite-dimensional diffusion models. By truncating the representation in the spectral domain we ensure our resulting model defines valid stochastic processes, thereby naturally satisfying consistency and exchangeability criteria. Projecting our spectral diffusion models back to the original input space, we show that for any given marginals our approach corresponds to a diffusion model with correlated noise, with explicit covariance matrix given by the kernel. We demonstrate our method's effectiveness for modelling various multimodal datasets as well as conditional sampling by amortising our models with respect to a context set.

A theoretical understanding of Transformers is crucial to better understand the capacities and limitations of large language models (LLMs). There is much work analyzing the expressivity of attention-based models. By proposing handcrafted weights or using computational complexity arguments, a large amount of past theoretical works have sought to characterize which tasks are and which are not in the hypothesis class of Transformer models. However, little work investigates the learnability of such solutions. In this work, we make progress towards this goal. Inspired by recent loss landscape analysis work, we propose preliminary sample complexity bounds for learning C-RASP constructions with Transformers.

Vision-language agents (VLAs) are increasingly used to interpret complex driving scenes and support safety-critical reasoning. This report presents the CVPR 2026@AdvML Workshop Challenge on adversarial multimodal attacks against autonomous-driving VLAs. Built on DriveLM-style multi-view visual question answering, the challenge represents each scene with six synchronized camera images and a structured collection of driving-related question-answer pairs. Participants generate adversarial images and suffix-only textual perturbations that induce model responses to deviate from reference answers while preserving image fidelity and limiting textual cost. The competition comprises two phases, with Phase II adding a hidden black-box model to assess transferability. We describe the task design, submission rules, evaluation protocol, and leaderboard results, and then examine five leading submissions for which technical reports were available. Across these reports, several recurring patterns emerge: image-side attacks are favored by the suffix penalty; scene-level, multi-view optimization is more effective than treating views in isolation; QA types and graph structure provide useful priors for allocating attack budget; feature-space objectives can improve black-box transfer; and typographic content embedded in camera images exposes a persistent vulnerability in driving VLAs. These findings provide a practical reference for future robustness evaluation and defense design in multimodal autonomous-driving systems.

The creation of digital collections involves not only the digitisation of content, but also the creation of catalogue records for it. This often-overlooked task requires slow and costly expert manual work. In this project, we have evaluated the application of AI models to this task, comparing different implementations and models. This work includes a qualitative and quantitative evaluation of the experiments carried out, as well as recommendations on the use of AI models that go beyond the specific use case.

Large language models (LLMs) are rapidly shifting toward agents that solve tasks through diverse interfaces, including web and graphical user interfaces (GUIs). Among these, the terminal command line provides a text-based, general-purpose interface, covering tasks from system operations to data science and machine learning. However, scaling terminal-agent training remains challenging, as it requires diverse and coherent task instructions, executable environments, and reliable verification, while lacking naturally grounded supervision data. In this work, we propose SETA, a scalable framework for generating verifiable terminal environments for reinforcement learning (RL). The framework consists of two pipelines sharing a unified verification mechanism: SETA-Synth converts diverse sources into standardized RL environments, and SETA-Evol further expands from existing environments with adaptive control of difficulty and diversity. Together, we construct and release SETA-Env, the largest open-source verifiable terminal RL dataset to date, containing over 4,500 environments. We evaluate our dataset by training Qwen3-8B with GRPO on SETA-Env, achieving 12% pass rate on Terminal-Bench 2.0, the best reported result for an RL-trained model at the 8B scale. We further observe gains on DeepSeek-V4-Flash under the same terminal agent harness, with pass@1 on Terminal-Bench 2.0 improving from 40% to 43% and pass@5 improving from 54% to 58%. These results demonstrate that SETA- Env provides high-quality training environments for terminal agents and serves as a valuable resource for advancing research on terminal-based agent learning.

Reinforcement learning (RL) is commonly employed to enhance the performance of autonomous systems, including the Autonomous Internet of Things (AIoT). However, the trial-and-error nature of RL, when conducted in real-world environments, is costly and hazardous in some scenarios. Consequently, the majority of RL research is conducted in simulation. This reliance introduces challenges related to the Sim-to-Real transferability. Evaluating the Sim-to-Real algorithmic robustness and the Sim-to-Real gap is a critical prerequisite for research aimed at improving RL performance in the real world. Therefore, industries such as robotics have developed concurrent simulation and physical platforms to facilitate this research. However, a universal Sim-to-Real benchmark platform for AIoT does not currently exist. To address these concerns, we developed a real-world AIoT platform for studying RL in AIoT. On this platform, an agent deployed on an edge device plays video games on a separate host computer via a hardware-emulated keyboard, guided by vision input. This platform uses commercially available components costing less than USD 400, together with two computers. Because the system's objective is game score maximization, it inherently mitigates safety risks associated with real-world RL deployments. Experimental results show the simulation-trained agent suffers a 1160% performance degradation relative to the human-level performance after real-world deployment, indicating a significant Sim-to-Real gap. Direct real-world training using the deep Q-network (DQN) algorithm achieves approximately 38% of human-level performance after 10 million training steps, demonstrating the feasibility of RL under real-world conditions. These results suggest that the proposed Sim-to-Real benchmark platform provides a substantial foundation for qualitative and quantitative evaluations of RL in real-world AIoT systems.

Existing 3D scene-grounded Large Language Models (3D-LLMs) focus on answering questions grounded in simplified single-room 3D scenes, lacking the ability to reason over real-world household environments containing multiple interconnected rooms and diverse object categories. We introduce CAIRN, a topology-aware 3D-LLM for multi-room 3D scene understanding. CAIRN aligns transformer attention with scene hierarchy, giving the model explicit awareness of object-level relations and room-level connectivity. It enriches object tokens with room-local relational context via a graph neural network, introduces learned room tokens for room-level abstraction, and applies a hierarchical attention mask with geometric bias to route information according to scene topology. CAIRN is developed on CAIRN-MR, a benchmark we introduce on HM3D for multi-room 3D scene understanding, covering grounding, captioning, and four question-answering tasks that progressively evaluate from intra-room perception to cross-room reasoning. Experiments show that CAIRN outperforms prior 3D-LLMs by a large margin across all CAIRN-MR tasks while remaining competitive on five single-room benchmarks.

As AI systems become more capable, training procedures that optimize for downstream outcomes risk introducing implicit agency: goal-directed behavior that designers never specified. We present a formal safety argument for the Scientist AI (SAI) Predictor, trained to approximate the Bayesian posterior conditioned on a dataset of "epistemically contextualized" natural-language statements. We argue that such a Predictor can honestly predict agents, actions, and their consequences without itself being an agent that selects outputs to achieve goals. This rests on data representation and on the training procedure. Epistemic contextualization of text distinguishes latent factual claims from communication acts, so expressions of goals are treated as evidence to be explained rather than drives the model adopts. With a posterior-seeking training objective, this is intended to drive the Predictor toward calibrated, cautious predictions. Training proceeds so downstream effects of deploying a prediction never serve as a reward signal; any agency the system needs is supplied by explicit scaffolding constrained by guardrails. We prove that, under assumptions on the training dynamics and on the argued sparsity of dangerous Predictors, the probability that training produces a Predictor whose guarded deployment carries residual harm above a specified threshold is small: a dangerous Predictor would have to underestimate harm in a coordinated way across many queries while such coordinated patterns are rare under the initialization distribution and receive no direct training signal. Safety and accuracy are jointly supported in this framework, since the constraints that secure accuracy are the same ones that make coordinated deception costly. These guarantees against misalignment and agency arising from within the Predictor itself do not preclude the use of the Predictor as part of an agentic system.

Unified and scalable Transformers have recently achieved remarkable success in modeling diverse phenomena traditionally associated with computer graphics, such as 3D visual effects, rendering processes, and motion in videos. In this work, we take a step further by investigating whether modern Transformer techniques can tackle the challenging task of cloth simulation. To this end, we present ClothTransformer, a framework that reformulates cloth simulation as autoregressive sequence modeling in a learned latent space. Existing neural cloth simulators are largely specialized to single scenarios, intrinsically coupled to the mesh discretization, and lack robust collision handling. Our approach addresses these limitations through three contributions: (1) a unified Transformer architecture that handles diverse scenarios -- body-driven garments, robotic manipulation, and free-fall collisions -- under a single model and achieves approximately $4$--$9{\times}$ lower error than prior state-of-the-art methods across all scenarios; (2) a scalable latent-space formulation that compresses arbitrary-resolution meshes into a fixed-size set of latent tokens, making temporal dynamics computation independent of mesh resolution; and (3) a diverse-scenario high-fidelity penetration-free dataset of ${\sim}$493.4k frames spanning all three settings, which enables a differentiable Continuous Collision Detection (CCD) module to suppress penetration artifacts. Project Page: https://yucrazing.github.io/clothtransformer/

While large language models (LLMs) are trained to align with human values, their generations may still violate safety constraints. A growing line of work addresses this problem by modifying the model's sampling policy at decoding time using a safety reward. However, existing decoding-time steering methods often intervene unnecessarily, modifying generations that would have been safe under the base model. Such unnecessary interventions are undesirable, as they can distort key properties of the base model such as helpfulness, fluency, style, and coherence. We propose a new test-time steering method designed to reduce such unnecessary interventions while improving the safety of unsafe responses. Our approach filters tokens using a value-based safety criterion and provides an explicit bound on the probability of false interventions. A single threshold hyperparameter controls this bound, allowing practitioners to trade off higher rates of unnecessary intervention for better output safety. Across multiple datasets and experiments, we show that our value-filtered decoding method outperforms existing baselines, achieving better trade-offs between safety, helpfulness, and similarity to the base model.

Large-scale video diffusion models achieve impressive visual quality, yet often fail to preserve geometric consistency. Prior approaches improve consistency either by augmenting the generator with additional modules or applying geometry-aware alignment. However, architectural modifications can compromise the generalization of internet-scale pretrained models, while existing alignment methods are limited to static scenes and rely on RGB-space rewards that require repeated VAE decoding, incurring substantial compute overhead and failing to generalize to highly dynamic real-world scenes. To preserve the pretrained capacity while improving geometric consistency, we propose VGGRPO (Visual Geometry GRPO), a latent geometry-guided framework for geometry-aware video post-training. VGGRPO introduces a Latent Geometry Model (LGM) that stitches video diffusion latents to geometry foundation models, enabling direct decoding of scene geometry from the latent space. By constructing LGM from a geometry model with 4D reconstruction capability, VGGRPO naturally extends to dynamic scenes, overcoming the static-scene limitations of prior methods. Building on this, we perform latent-space Group Relative Policy Optimization with two complementary rewards: a camera motion smoothness reward that penalizes jittery trajectories, and a geometry reprojection consistency reward that enforces cross-view geometric coherence. Experiments on both static and dynamic benchmarks show that VGGRPO improves camera stability, geometry consistency, and overall quality while eliminating costly VAE decoding, making latent-space geometry-guided reinforcement an efficient and flexible approach to world-consistent video generation.