Papers

This report examines and compares some of the many fine tuning and prompting methods existing, applying them within the domain of autonomous driving. The idea is to compare these methods by adapting a Large Language Model (LLM) on a video dataset. LLM's have become extremely good at achieving a good understanding of different forms of data and this study aims to induce a low dimensional understanding of driving situations into our primary test model SpaceTimeGPT. Experiments on BDD-X (Berkeley DeepDrive eXplanation) dataset demonstrate good performance of the full fine tuning framework on some automatic metrics, and in some metrics, it even surpasses the baseline. We also try Low-Rank Adaptation (LoRA) and prompt engineering on VideoLLaVA model and discuss its limitations.

The EU AI Act (Art. 27) requires deployers of high-risk AI systems to conduct Fundamental Rights Impact Assessments (FRIAs) before deployment, yet the evidence needed for credible assessments is fragmented across incompatible incident repositories, risk vocabularies, and legal texts. We present a reusable Semantic Web-based framework that consolidates this evidence for two high-risk public sector categories: employment and worker management (Annex III(4)) and access to essential public services (Annex III(5)(a)). A curated 150-record corpus is annotated along four axes using keyword, LLM, and hybrid methods and serialised as a SPARQL-queryable knowledge graph of 1,351 RDF triples. Five FRIA demonstration scenarios surface 103 records (68.7% coverage). Evaluation against a 69-record gold standard reveals that LLM-assisted classification of the employment domain achieves only $\kappa = 0.045$, a cautionary result for automated fairness-related evidence retrieval in this domain. All artefacts are released openly to support adoption by regulators, national authorities, and SMEs.

During reinforcement learning with verifiable rewards (RLVR), large language models (LLMs) can exploit loopholes in their environments to obtain high rewards without improving the intended capabilities, i.e., reward hacking. Despite its risks to training efficiency and safety, monitoring and mitigating reward hacking during training remain challenging, which is limited by a lack of testbeds that reproduce hacking and reliably identify it. We introduce CATCH, a controllable testbed for studying reward hacking in coding RL. CATCH deliberately exposes environmental loopholes and provides execution-based gold labels by comparing success under a vulnerable evaluator with task correctness under an independent audit. It also can control the model's initial hacking tendency through supervised fine-tuning data mixtures and the difficulty of earning rewards through reward designing, enabling systematic comparisons of hacking dynamics and interventions. Experiments show that CATCH can produce diverse RL training trajectories with clear reward hacking, and analyses demonstrate that both initial models and reward difficulties shape the emergence of reward hacking. We further evaluate the effectiveness of different reward hacking detection and mitigation methods. A key finding is that a chain-of-thought monitor initially suppresses hacking, but this protection erodes as the policy model learn to mislead the monitor with code comments. This highlights the need to evaluate hacking mitigations throughout training with CATCH. The source code and resources are publicly released at https://github.com/THUAIS-Lab/CATCH.

We study the efficiency of auctions with budget-constrained bidders and derive tight price of anarchy (PoA) bounds for coarse correlated equilibria (CCE). To this end, we develop a general framework that reduces proving PoA guarantees to finding feasible solutions to the dual of a linear program. The LP is formulated over a small number of per-bidder equilibrium statistics, incorporates constraints implied by the auction setting and equilibrium behavior, and captures a novel smoothness notion tailored to budget-constrained environments; we provide a reduction from classical smoothness to this new notion. Our approach enables simple and tight PoA analyses across a broad range of auction formats. Using our framework, we obtain tight liquid welfare guarantees for simultaneous first-price, second-price, and all-pay auctions, simultaneous auctions with restricted uniform bidding interfaces, discriminatory and uniform price multi-unit auctions, and generalized first-price position auctions. Beyond budget-constrained settings, we derive new lower bounds for the budget-free uniform price auction and generalized second-price auction; in particular, our lower bound for the uniform price auction matches the upper bound of $3.146$ by de Keijzer et al. (2013).

We study the problem of estimating the guidance that steers the distribution learned by a diffusion generative model toward a tilted target $q_0 \propto w\,p_0$ at inference time. Relying on the stochastic optimal control approach, we observe that the exact drift correction is the gradient of the logarithm of Doob's $h$-function, and we study the problem of estimating it from a sample. In the present paper, we assume that the score of the pretrained model is available, that the tilting weight is bounded and positive, and that the reference distribution has a bounded support, no smoothness of the weight is required. Introducing a penalized least-squares risk in which the penalty is the residual of the space-time harmonicity equation satisfied by the $h$-function, measured in a dual Sobolev norm, we derive high-probability bounds on the squared error of the resulting guidance estimate. Since the penalty vanishes at the target, the estimator is free of regularization bias, and in favourable scenarios its rate of convergence is faster than the minimax rate of estimating first-order derivatives of a smooth regression function. Assuming that $w$ is bounded and positive with $\mathbb{E}_{p_0}[w^{-\mathrm{s}}] < \infty$ for some $\mathrm{s} \in (0,\infty]$, and that the reference data are compactly supported, we prove that the guidance is estimable in squared $L^2$ at rate $\varepsilon_n^{\mathrm{s}/(\mathrm{s}+4)}$, where $\varepsilon_n = n^{-2(\beta-1)/(2(\beta-1)+d)}.$ We also transfer the obtained bounds to the total variation distance between the marginals of the estimated and the exactly guided samplers, and illustrate the performance of the suggested approach with numerical experiments.

eBPF observability of non-standard execution environments (NEEs) like TEEs or LibOSes is hindered by their unconventional exception-handling and memory-access mechanisms that limit standard Linux tooling. This work introduces two mechanisms that enable eBPF-based observability for NEEs: (1) kernel memory extensions for safely accessing NEE memory from eBPF programs, and (2) a lightweight flexible probe performance measurement unit (LWFP PMU) that provides flexible and generic probing, for NEEs, through the following LWFP: simple (SLWFP), enclave (ELWFP) and extended (ExLWFP) probes. We demonstrate the practicality of these extensions by developing tooling for tracing, stack sampling with Flame Graphs, dynamic instrumentation, timing analysis, and USDT support for Intel SGX enclaves and LibOSes. Performance measurements show that SLWFP probes achieve a latency of 394 ns, outperforming uprobes, which exhibit 25% higher latency, while ELWFP probes incur a latency ~2.8 microseconds, which is practical for enclave observability. The addition of SLWFP introduces negligible overhead to existing uprobe performance. Taken together, this work lays the foundation for closing the long-standing gap between NEE tooling and Linux observability tooling by enabling generic and reusable eBPF tooling for diverse NEE hardware and software architectures.

The presence of lens flares in images can significantly reduce the quality of downstream application results for tasks such as 3D scene reconstruction. This is because lens flares are a property of the camera imaging system, and not a part of the underlying scene being modeled. There are previous methods that tackle the removal of small flares focused around a light source. However, existing methods struggle with large flares, such as those that fill the entire image. In this work, we compile a novel dataset for large-flare removal, combining publicly available real-world data with a procedural generation pipeline. We fine-tune a diffusion-based model on our dataset to remove complex, large lens flares. On the other hand, lens flares remain effective artistic tools, widely used in the media. While there are ways to simulate 2D flares, representing and reconstructing lens flares consistently across multiple views has not yet been explored. To achieve this, we introduce a flare representation model that leverages the symmetry of lens flares about the camera's principal point. We propose a computational pipeline to jointly optimize this flare model and a Gaussian splatting model (3DGS). This enables the decomposition of a 3D scene into lens flares and the scene itself, using our flare-removal model. Because the reconstructed flare is explicit and re-renderable, it can be edited and transferred to novel images and new 3D scenes. We evaluate removal on an established benchmark and a new one for large reflective flares, quantify the flare/scene decomposition directly, and show that the pipeline is robust to errors in automatic light-source localization.

Efficient action generation in vision-language-action (VLA) models requires capturing both coarse action structure and fine-grained details. Discrete action tokens provide compact structural representations but sacrifice precision, while continuous action tokens offer high precision but often require multiple denoising steps. We introduce Discrete Forcing, a flow-matching framework that combines these representations through an explicit coarse-to-fine generation process. It first predicts discrete action tokens to establish a coarse action structure, then uses them to guide continuous action refinement. The discrete and continuous components share a common diffusion transformer backbone with specialized branches, maintaining a parameter count comparable to a conventional single-branch model while requiring only one forward pass per branch. Extensive evaluations across multiple benchmarks demonstrate improved performance and faster inference over a parameter-matched continuous action expert, with consistent performance gains as model capacity increases. Real-world experiments further demonstrate improvements on high-precision and dynamic manipulation tasks.

Casting Bayesian inference as a neural network optimization problem targeting an amortized posterior is attractive, as it extends to otherwise intractable statistical models and offers near instantaneous inference for new datasets after prepaying the training cost. Although theory guarantees faithfulness under ideal convergence, practical amortized inference still requires iterating over architectures and optimization choices and ultimately “satisficing” under finite simulation, compute, and time budgets. Even the best-performing solution may thus retain avoidable representation gaps that typically require problem-specific fixes. Here, we propose a generic alternative which improves training dynamics with auxiliary guidance losses applied to internal representations. Specifically, we show how such guidance leads to faster convergence when training data is abundant and to better performance when it is scarce. We formalize representation gaps as getting stuck in a local optimum at the information bottleneck between the parts of the network tasked with feature learning and those tasked with conditional distribution learning, and offer a generic diagnostic to separate summary failures from inference failures. Finally, we demonstrate that auxiliary supervision improves convergence speed and accuracy on a range of challenging real-world inference problems.

Causal foundation models (CFMs) amortise causal inference over priors of synthetic structural causal models (SCMs), predicting the effect of an experiment on a specific variable. However, observational data alone may leave multiple causal models compatible with available evidence, while experimental data with interventions on exactly the variable of interest might be unavailable. This work studies CFMs as a method to combine finite observational and surrogate-interventional datasets in order to predict a target conditional interventional distribution (CID) more accurately than with observational data alone. We first formalise the conceptual benefits of surrogate experiments. Building on this analysis, we introduce Foundation Models for Causal Inference from Diverse Experimental Regimes (CIDER-FM), a causal foundation model that uses an intervention-aware representation and hierarchical three-axis attention to exchange information across variables, samples, and experimental regimes. We evaluate CIDER-FM against a wide range of baselines across diverse synthetic graph and mechanism families, as well as on both simulated and real-world data from Causal Chambers. Our results demonstrate strong CID prediction performance and show that incorporating experimental context can improve predictions over observational data alone.

Effective human-AI collaboration requires partners to establish references through interaction, which becomes fragile when descriptions are ambiguous, similar referents compete, or partners see different things. We study referential uncertainty - uncertainty over which candidate object a description refers to - in a collaborative puzzle task where a human Helper instructs an AI Worker to place pieces. The Worker must identify and communicate its uncertainty, and the Helper must recognize and act on it. We show that a separately elicited belief distribution over candidate pieces is better calibrated (ECE 0.15) and better discriminates correct from incorrect placements (AUROC 0.65) than raw action-token probabilities, which are severely overconfident (0.97 mean confidence, ECE 0.44). Across three frontier vision-language models (GPT-4.1, GPT-5, GPT-5.5), this elicited uncertainty rises predictably with instruction vagueness, but not with competing referents in context, even when those increase errors. The models seldom externalize it, asking for clarification on only 3.5-16.7% of turns. In a controlled human study (N=210), participants given only the Worker's default message accept 78% of wrong placements and cannot tell right from wrong (AUC 0.50). Precise descriptions and, especially, well-targeted hedges cut wrong-move acceptance to 36% while largely preserving correct-move acceptance, compensating for missing shared awareness such as not seeing the Worker's action. But this benefit depends on targeting: a deployable hedge derived from the model's own belief entropy inherits that signal's weakness and can do more harm than good. Externalized uncertainty helps a human partner only when it is accurately targeted.

We give a full list of translational nets which flex within their class of discrete surfaces of translation, by reducing the classification problem to the one of flexible complete bipartite frameworks on the sphere, for which the solution is known. We also obtained two novel classes which correspond to Bottema's spherical 16-bar mechanisms and the constant diagonal angle frameworks. Based on an algorithm for the construction of all flexible translational nets, we also discuss flexible translational tubes and toroids. Furthermore, we present novel results for both topologies which are implied by Bottema's spherical 16-bar mechanisms.

Vision-language-action (VLA) models bridge multimodal understanding and robotic control, advancing the dominant paradigm for embodied intelligence. However, most existing models rely on large Transformers, whose latency and energy costs hinder deployment on resource-constrained platforms. Through sparse event-driven computation, spiking neural networks offer a promising paradigm for high-performance and energy-efficient computing. Here, we propose the first Spike-driven VLA framework enabling end-to-end direct training for robotic manipulation, which mainly comprises three core components. First, we develop spiking visual and instruction encoders for multimodal perception, encoding visual observations and language instructions into sparse, reliable spike representations for subsequent cross-modal fusion. Then, we introduce Multi-Winner Spike Fusion for instruction-guided scene understanding, using bidirectional top-$k$ winner-take-all spike routing to suppress background interference and yield fused memory. Finally, we propose a Spike Action Chunking Transformer that incorporates spiking cross-attention over the fused memory and the current robot state, enabling efficient end-to-end generation of continuous action chunks for robotic control. Extensive experiments on LIBERO and Meta-World demonstrate that Spike-driven VLA achieves competitive performance with fewer parameters and lower estimated inference energy than conventional VLA models. This work establishes a foundational framework for neuromorphic VLA modeling, paving the way for future advances in resource-efficient embodied intelligence.

Despite being an emerging credible means of using artificial intelligence (AI) for improving disease surveillance via early detection of outbreaks, epidemics prediction and evidence-based decision-making, there continue to be challenges in the use of AI tools in many low- and middle-income countries (LMICs) due to various factors including fragmented health information system, digital inequality, governance problems and institutional incapability. Most of the research conducted so far has concentrated on the efficacy and accuracy of AI in terms of predicting outbreaks, with little focus on other conditions necessary for its deployment. This study adopts a qualitative problem-discovery research design, integrating thematic analysis with philosophical analysis to examine the structural and normative barriers surrounding AI implementation in disease surveillance. The analysis identifies four interrelated dimensions of implementation readiness: epistemic adequacy, distributive justice, ethics of governance, and institutional legitimacy. These dimensions provide a framework for understanding how limitations in knowledge integration, unequal digital infrastructure, privacy and accountability concerns, and deficits in institutional and public trust can constrain the practical adoption of AI-enabled surveillance. Rather than proposing another predictive model, this study develops a theoretical framework that conceptualizes AI implementation in disease surveillance as simultaneously a socio-technical and normative process. The framework provides a foundation for subsequent empirical investigation and offers a structured perspective for designing more context-sensitive, ethically grounded, and institutionally sustainable AI-enabled disease surveillance systems in resource-constrained healthcare settings.

The small-sample learning problem remains a fundamental challenge in machine learning because limited training data lead to unstable model estimation and generalization. Structural Risk Minimization (SRM) has long been regarded as a principled solution under the classical i.i.d. assumption. However, domain generalization (DG) violates this assumption, leaving the theoretical role of SRM in DG largely unexplored. To bridge this gap, we establish the first theoretical guarantees for SRM in DG under mild assumptions. Specifically, based on the concept of stability, we derive learning consistency and generalization error bounds and prove that these bounds become tight when the hypotheses satisfy the stability condition. Building upon this, under a specific hypothesis space assumption, we establish stability, learning, and generalization bounds for SRM. We further discuss the applicability of these bounds to deep learning. This work establishes theoretical foundations for SRM under distribution shifts and sheds light on the design of robust DG algorithms in small-sample scenarios.

The geometrically decomposed bases of the polynomial differential form spaces $\mathcal{P}_r\Lambda^1$ and $\mathcal{P}_r^-\Lambda^1$ on a simplex are products of a barycentric scalar polynomial and a directional or Whitney 1-form. These bases depend on an ordering of the simplex vertices. On unstructured meshes, neighbouring cells need not agree on this ordering, making it non-trivial to achieve conformity in finite element software. Existing strategies compute degree-of-freedom transformation matrices numerically, or impose a global vertex ordering through face-bubble spaces or by preprocessing the mesh. We consider the standard basis of the $\mathcal{P}_r^-\Lambda^1$ (trimmed) space and introduce a new basis for the $\mathcal{P}_r\Lambda^1$ (full) space, for arbitrary polynomial order $r \geq 1$ and dimension $D \geq 2$, and prescribe their basis polynomials as shape functions. We show that the pullback of shape functions under the change of coordinates induced by an arbitrary vertex relabelling $\pi$ in the symmetric group $S_{D+1}$ admits a closed-form, combinatorial expression. For most basis functions this pullback is a single basis function of the relabelled ordering, up to sign. On an explicitly characterised "filter-hit" set, this pullback is a signed sum of at most $D$ (full space) or exactly two (trimmed space) relabelled basis functions. All coefficients are in $\{-1, +1\}$ for all $r$, $D$ and $\pi$. The inverse transformation is obtained by computing the formulas at $\pi^{-1}$, so no numerical inversion of a change-of-basis matrix is needed. Conforming assembly and evaluation on simplicial meshes with arbitrary vertex orderings reduce to index manipulation. Our results are compared with the study of relabelling-invariant bases of Berchenko-Kogan and Licht, and verified in an open-source Julia implementation in the Gridap.jl library.

Let $W_s=K_1+C_{s-1}$ denote the wheel on $s$ vertices. We give structural proofs that $R(C_{14},W_{11})=27$ and $R(C_{15},W_{11})=29$. Together with the theorem of Chen et al. for $n\ge16$, these equalities give $R(C_n,W_{11})=2n-1$ for every $n\ge14$. The two boundary values were included in an earlier survey announcement. We also give structural proofs of $R(C_8,W_7)=15$, $R(C_9,W_7)=17$, and $R(C_8,W_9)=15$. The common starting point is a Hamilton-connected core lemma. For the eleven-vertex wheel, bounds on vertex connectivity and on the matching number of a bipartite graph associated with a local cycle yield a vertex cut of order nine. Paths with prescribed endpoints then rule out every possible pair of orders of the two remaining vertex sets. For the smaller wheels, we use the structure of critical cycle colorings and local cycle-shortening arguments. We also give complete structural classifications of the $(C_8,C_6)$- and $(C_9,C_6)$-critical colorings, recovering the previously reported counts 24 and 26. All proofs are combinatorial and use no exhaustive graph enumeration.

General-purpose agents can plan, use tools, and revise their behavior from feedback, but it remains unclear whether these capabilities transfer from digital environments to embodied manipulation. To investigate this question, we introduce LIBERO-Agent, an agent-native benchmark for evaluating these agents in robot manipulation tasks. Rather than asking agents to submit task-level Python control programs or operate through high-level robot skills, LIBERO-Agent provides an interactive robotic environment where agents can select which observations to inspect, process them with their own tools, and issue native action commands. LIBERO-Agent integrates 200 tasks into a common interaction framework and provides a 30-task primary suite that separates perception, short-horizon execution, and long-horizon composition. Results reveal a pronounced reliability gap: while agents perform well on perception and easy short-horizon tasks, their performance degrades substantially on hard short-horizon and long-horizon tasks. Richer observations improve short-horizon manipulation, while demonstration benefits depend on the agent and format. Among these agents, GPT-6 Astra achieves the strongest overall performance. Further analysis shows its major advantage lies in mechanism interaction, especially when sustained physical contact is needed, while its remaining failures stem from cross-stage interference and geometric errors.

We present PartiCam, a training-free Particle filtering rooted method for improved Camera controlled video generation. Generating videos that follow a precisely specified camera trajectory remains challenging for large video diffusion models. Training-free approaches are backbone-agnostic and avoid the need to construct large camera-annotated datasets by steering pretrained models toward the desired camera motion at test time. This enables the generation of camera-controlled video data that can subsequently be used to train camera-conditioned video diffusion models. Existing sampling-based guidance approaches often suffer from unstable trajectories: they either explore too broadly and fail to respect the target camera motion or collapse early and lose visual diversity over time. We introduce a global-local refinement framework for diffusion reward guidance, enabling accurate and consistent camera control during video generation. Our method builds on Sequential Monte-Carlo (SMC) guidance, but introduces a local refinement stage based on particle filtered resampling. Experiments show large improvements in camera trajectory adherence, reduced drift, and better visual quality, without requiring model retraining.

Randomized smoothing certifies the probability of a fixed output event as the center of Gaussian noise moves. Feasibility or confidence filtering reports label probabilities only among retained proposals, producing a ratio. Its numerator is a fixed Gaussian event mass, while its denominator is the probability of retention and can change with the center. Substituting this ratio into the ordinary smoothing formula can therefore certify a ball that contains a decision boundary. We separate the problem into a geometric question and a certification question. Geometry determines when conditioning preserves Gaussian comparisons. Convex retained sets preserve the full comparison, while general sets require geometric control of the retained law as the center moves. Without such control, conditional probabilities imply no positive universal radius. Joint retention-and-label probabilities always yield a valid certificate for the same filtered predictor. A uniform covariance bound transfers divergence certificates to the retained law and can yield larger radii even when the Gaussian event comparison fails. Both methods admit finite-sample bounds. For a learned image classifier with a training-selected nonconvex filter, conditional Rényi bounds certify more images than joint-mass bounds without additional model evaluations. A released confidence filter exhibits verified label changes inside radii obtained by conditional substitution. An application of adaptive Gaussian composition covers causal finite-horizon executions with history-dependent center shifts under a pathwise energy bound.

Typed decision models such as Jev offer an efficient alternative to generative LLMs in decision-making workflows by selecting directly from predefined options. When candidate sets contain no valid answer, TypeSafe recommends including an "other" or "none-of-the-above" option to enable rejection. In this report, however, we identify an arithmetic-dependent rejection bottleneck: Jev reliably selects correct numerical answers when available but frequently accepts incorrect alternatives when they are absent despite an explicit rejection option. On paired arithmetic problems, answer-present accuracy reaches 99%, while correct rejection falls to 7%. Moreover, this gap persists across numerical magnitudes, operation depths, contextual formulations, and rejection labels, and extends to scenarios such as time calculation and capacity rounding. Yet native Boolean verification achieves 99% exact-match accuracy on the same answer-absent arithmetic cases, showing that categorical rejection can fail even when the model successfully verifies candidate correctness. Finally, we show that a simple decision threshold selected on separate development problems raises arithmetic rejection accuracy from 7% to 79% while retaining 97% answer-present accuracy, substantially mitigating the failure without retraining or additional inference.

Interactive proof systems with a laconic prover, studied by Goldreich, Vadhan, and Wigderson (CC, 2002), capture problems verifiable with logarithmic prover communication in the classical setting. For two-message quantum analogs, even a single-bit prover response contains quantum statistical zero-knowledge ($\sf QSZK$), introduced by Watrous (FOCS 2002). However, restricting the verifier's question to classical public coins collapses the corresponding class to $\sf BQP$, as shown by Beigi, Shor, and Watrous (ToC, 2011). We further study two-message quantum interactive proof systems with a laconic prover. To this end, we introduce the class ${\sf QIP}_{\ell\text{-}{\rm bit}}(2)$, where $\ell$ is the length of the prover's response, and establish: 1. A natural complete characterization of ${\sf QIP}_{\ell\text{-}{\rm bit}}(2)$ by Multi-State Distinguishability. In particular, Quantum State Distinguishability (QSD) is ${\sf QIP}_{\rm bit}$-complete. Since QSD is $\sf QSZK$-hard, our result places ${\sf QIP}_{\ell\text{-}{\rm bit}}(2)$, for $\ell\geq 2$, in a landscape "just above" $\sf QSZK$. 2. Easy regimes for ${\sf QIP}_{\ell\text{-}{\rm bit}}(2)$ collapsing to $\sf QSZK$. We prove that QSD$[a,b]$ (and thus ${\sf QIP}_{\rm bit}[a,b]$) is in $\sf QSZK$ when $a(n)-b(n)\geq 1/O(\log n)$, and combine this with an answer compression from ${\sf QIP}_{\ell\text{-}{\rm bit}}[2,c,s]$ to ${\sf QIP}_{\rm bit}$ to obtain another easy regime when $2c>(1+2^{\ell/2})s$. Remarkably, our improved polarization applies to SD and $\sf SZK$, resolving an open problem in Sahai and Vadhan (JACM, 2003). 3. Quantum public coins also make the interaction useless: ${\sf qc}\text{-}{\sf QAM}[O(\sqrt{\log{n}})]$ with constant gap is in $\sf BQP$, where ${\sf qc}\text{-}{\sf QAM}[\ell]$ is a subclass of ${\sf QIP}_{\ell\text{-}{\rm bit}}(2)$ in which the verifier's question is exactly halves of EPR pairs.

Self-distillation with privileged context adapts a language model from demonstrations by letting the model, once conditioned on a reference response, teach its context-free copy token by token. Our taxonomy reveals existing methods differ along three entangled axes: (i) the rollout source (student or teacher), (ii) the teacher coupling (frozen, or an exponential moving average of the student at some coupling rate) and (iii) the KL direction (reverse or forward), yet these axes are usually studied in fixed combinations and have led to conflicting conclusions. We formalize a unifying framework to encompass all self-distillation methods vs classic supervised fine-tuning: we train every combination of the three axes, on Qwen2.5-7B and Ministral-3-3B across ordinary and contradictory tasks, totaling 1,200 adaptation runs, to systematically investigate the impact of the above axes. We propose a controlled model of the same objective to explain the resulting acquisition-retention trade-offs. We find that (i) the rollout source matters mostly where the task contradicts the pretrained behavior: there teacher rollouts raise acquisition well above what student rollouts achieve, with almost no change in retention; (ii) the teacher coupling changes acquisition most, on every task: acquisition rises with the coupling rate, then falls past a task-specific rate; (iii) switching the KL direction costs retention in one model but not the other so which axis to tune first depends on the model. The controlled model reproduces the three trends.

Matching the same vehicle across front and rear cameras is difficult because the cameras do not share a view and the vehicle's appearance changes substantially. We introduce Front2Back-ReID, a benchmark of 500 manually verified vehicle handovers from 20 recording sequences in South Africa. Each example asks a model to match a vehicle highlighted in a front-camera image to the same vehicle among at least three candidates in a later rear-camera image. We evaluate seven zero-shot vision-language models, four image-retrieval baselines, and 25 human participants. Models are tested using full front RGB images, cropped target vehicles, and binary silhouettes. The strongest VLM achieved 76.6 percent Rank-1 accuracy on target crops, compared with 74.0 percent for the frozen SigLIP2 baseline; this difference was not statistically clear. Human participants achieved 94.0 percent accuracy with full images and 92.2 percent with target crops. Under our evaluation setup, enabling reasoning improved accuracy across all three input conditions for every model evaluated in both modes. We also found that VLMs generally performed worse on full scenes than on target crops. These results show that general-purpose VLMs do not yet consistently outperform strong visual retrieval for front-to-rear vehicle matching, while humans remain substantially more reliable.

Recent advances have enabled unified omni-modal models in understanding audio, vision, and language. However, existing benchmarks, training data, and learning methods largely treat the modalities independently, leaving the capability of audio-visual joint reasoning poorly evaluated and insufficiently elicited. We address this gap with a benchmark, data engine, and learning method. First, we introduce OmniReasoningBench, a benchmark where both audio and visual evidence are indispensable. It comprises 1,150 multiple-choice and open-ended questions across two tasks, reasoning over video and reasoning beyond video. Second, we develop a data engine OmniQA. It automatically constructs evidence-grounded QA pairs that explicitly necessitate audio-visual joint reasoning, together with time-stamped clue chains that guide the annotation of thinking process. Besides our benchmark, this engine produces training data OmniReasoning-SFT-112K and OmniReasoning-RL-19K. Finally, we propose an on-policy self-distillation method Modality-Factored Self-Distillation (MFSD). It evaluates each sampled response under modality-specific clue contexts, disentangling the contributions of individual clues and their cross-modal interactions for token-level credit assignment. With our training data and learning method, our model OmniReasoning-30B-A3B achieves 50.0% on OmniVideoBench and 42.5% on OmniReasoningBench, improving the base model Qwen3-Omni-30B-A3B-Thinking by 12.8 and 9.3 percentage points, respectively. Moreover, it delivers substantial gains on general and long-video benchmarks, including Video-MME-v2. We hope our work offers a solid step for facilitating future research in omni-modal joint reasoning.

Sample-level reliability heterogeneity is common in deep time series learning. Standard training pipelines apply a uniform regularization setting to all samples, which can under-regularize corrupted samples and over-restrict clean samples. Common robustness approaches filter observations in data space or impose priors on latent representations. We propose Capacity-Centric Modulation (CCM) as a complementary, sample-adaptive regularization principle. Under this principle, we introduce SACM (Sample-Adaptive Capacity Modulation), a task-agnostic framework that exploits spectral sparsity to assign sample-wise dropout probabilities along internal activation paths. SACM integrates into existing backbones without architectural redesign and preserves the deterministic inference pipeline. Across 301 real-world dataset-backbone pairs covering 9 forecasting, 32 classification, and 4 anomaly-detection datasets, SACM reduces forecasting MSE by 6.7% on average and improves classification accuracy and point-adjusted F1 by 3.04% and 17.05%, respectively, relative to unmodified backbones, with zero test-time overhead.

Flow matching generates samples by gradually transforming noise into data. In practice, using a finite number of sampling steps introduces a numerical error that depends on the chosen schedule. We study this dependence for Gaussian targets and the explicit midpoint sampling method, using the exact flow field. We measure sampling error by the squared Wasserstein distance between the target distribution and the final distribution produced by the midpoint sampler. We show that the standard conditional optimal transport (CondOT) schedule cancels the leading midpoint error and improves the general convergence bound, even when the sampling steps are unequally spaced. On a uniform grid of $S$ sampling steps, we fix the signal schedule at $\alpha_t=t$ and prove the existence of scalar noise schedules $\beta_t$ that approach the CondOT noise schedule $1-t$ at rate $1/S$ and yield exact Gaussian sampling for every sufficiently large $S$. Controlled Gaussian experiments illustrate the convergence rates and exact calibration.

Estimating accident mechanics from real-world crashes is important for vehicle-safety analysis, injury modeling, crash-severity prediction, and operational workflows such as insurance claim triage. In standard crash records, key metadata such as impact configuration, principal direction of force, and change in velocity ($\Delta V$) may be missing, delayed, or corrupted, while post-crash photographs are widely available and contain rich visual evidence of deformation. We study how much crash-mechanics information can be recovered directly from vehicle photos when structured signals are absent. We formulate crash understanding as supervised prediction from per-case multi-view photo sets. Targets include six Collision Deformation Classification (CDC) descriptors and the longitudinal and lateral components of reconstructed $\Delta V$. Each photo is encoded by a shared visual backbone, and the resulting view-level features are fused into a case-level representation from which target-specific heads predict crash descriptors. Using 15.2k training cases from the Crash Investigation Sampling System, drawn from about 1.5M photos before filtering, together with 1.15k validation and 1.15k test cases, we define an evaluation protocol for vision-based crash descriptor estimation from incomplete multi-view evidence. Post-crash imagery alone provides usable signal for several non-trivial crash-mechanics descriptors, while weakly observable and long-tailed targets remain challenging. Within the compared training regimes, the selected joint-training recipe reduces mean absolute angular error for principal direction of force from 20.1 to 14.05 degrees and longitudinal $\Delta V$ MAE from 8.04 to 7.45 km/h. Our work provides a reference point for future multimodal fusion with structured crash metadata.

Longitudinal clinical cohorts are multimodal, irregularly sampled and pervasively incomplete: in ADNI, positron emission tomography and cerebrospinal fluid assays are absent from roughly half of all visits. Linear state-space models handle irregular sampling gracefully but treat a missing modality by masking the input, leaving the transition operator untouched. We prove that this is a representational limitation: the latent state of any linear state-space layer whose transition operator does not depend on the availability pattern is an additive function of the availability indicators, so no such layer can represent an interaction between two modalities being jointly present or jointly absent. We propose CAMOS, which gives each modality a bank of second-order oscillators coupled through a matrix that sits inside the differential equation and is gated by availability, so the transition operator itself becomes a function of which measurements were taken. Coupling invalidates the analysis of uncoupled oscillatory models, and we restore it: a per-channel Gershgorin budget makes the effective stiffness positive definite uniformly over all $2^M$ availability patterns and all gaps, an energy argument charges amplification to availability transitions rather than sequence length, and a channel factorization preserves exact associative parallel scans. On ADNI, CAMOS outperforms uncoupled oscillatory state-space models and clinical fusion models on same-visit staging, landmark prediction and longitudinal forecasting, and under zero-shot transfer to OASIS-3 it is the only model that avoids collapse to the majority class.

Ownership establishes rights over the use, control, and transfer of objects. Understanding these relations is essential for AI systems to interact appropriately with people and their resources. Yet how large language models (LLMs) attribute ownership under competing claims remains unclear. We introduce the Competing Ownership Attribution Task (COAT), comprising 42 scenarios, and compare ownership allocations from 24 LLM configurations with those of 108 human participants. Overall, human-model similarity is close to human-human similarity, but models show greater homogeneity in their ownership judgments. Within individual answers, models also divide ownership more evenly among claimants than humans do. Pooling responses across model configurations reveals more scenarios with a shared judgment and fewer with distinct viewpoint groups than in humans. When humans form distinct groups, models may converge on one viewpoint or between competing viewpoints. Further comparisons reveal different contextual sensitivities. As material value increases across scenarios, allocations to creators decline less sharply in models than in humans. Across scenarios differing in public recognition of later holders as owners, allocations to these holders increase in models but decrease slightly in humans. Together, these findings suggest that the evaluated LLM responses do not fully capture the diversity of participants' ownership judgments or how those judgments vary across situations. Developing socially capable AI therefore requires moving beyond overall similarity to capture the diversity and context dependence of human judgments.