Science & Mathematics

Reasoning models often run into problems without noticing them, even when symptoms are visible in their own text. In humans, metacognition separates monitoring from control: fast, intuitive signals keep watch while slow, deliberate reasoning unfolds and help decide whether to continue, change course, or stop. System One (S1) decision models, such as Jev, are a natural candidate for this role: they return a calibrated probability for a natural-language question about a text in one fast call, yet they are used only as gates before reasoning. We introduce SOMA (System One Metacognitive Awareness), in which an S1 model continuously monitors a System Two (S2) reasoning model and intervenes when needed. SOMA has three layers: a scenario layer states each objective in natural language with predefined actions, an awareness layer lets S1 judge whether the described behavior appears, and a control layer accumulates these judgments, first reminds S2, and takes over only if the behavior persists. Sharing one S1 model and one controller across tasks, without training, SOMA addresses two different failures on two reasoning models: on overthinking, it cuts reasoning tokens by up to 36\% while improving average accuracy, and on unanswerable questions, it raises abstention from 70.5\% to 87.2\%.

Recent work in tabular machine learning has emphasized the importance of diverse and realistic evaluation settings. The niche of large-scale (1M+) data, however, remains under-explored. To study training-data scaling, we assemble eight real-world tabular datasets with up to around ten million training examples, retaining the natural distribution shifts of the tasks and using random-split evaluation as a control. We find that GBDTs and MLP-based models show broadly similar scaling trends, with gains from both additional training data and algorithmic improvements. Tabular foundation models show more varied scaling behavior, with recent versions benefiting more consistently from larger contexts even under distribution shift. However, scaling training data under distribution shift introduces new challenges: in-distribution and out-of-distribution performance mostly improve together at smaller scales but can diverge at larger scales. We examine this through MLP capacity sweeps and comparisons of the validation distributions used for tuning and early stopping in MLPs and GBDTs. We also compare common practices of using validation data, contrasting its use for model selection in MLP pipelines with using it for in-context fitting in tabular foundation model pipelines. Under stronger shifts, MLPs selected using validation representative of the evaluation conditions generally maintain or improve performance as training data grows. In contrast, including such examples in context improves foundation-model performance, but does not prevent deterioration with increasing training size in our experiments.

Every continuous piecewise linear (CPWL) function is a finite signed sum of maxima of affine functions, making maximum representation a central problem in the study of the expressiveness of ReLU networks. A now-disproved conjecture predicted logarithmic depth for maxima. Inspired by the proof by Brunck et al. (2026) that two hidden layers suffice for general CPWL functions, we give the Caterpillar Net, a nonrecursive construction for the maximum of $N$ inputs with two hidden layers of widths $O(N^2)$ and $2^{O(N\log N)}$. The construction has a direct geometric interpretation.

Financial institutions increasingly deploy large language models (LLMs) in multilingual markets, raising a practical question: can pipelines developed and validated in English be reused across languages, or must they be rebuilt per locale? We test cross-lingual reusability along three dimensions--accuracy, reasoning-effort transfer, and behavioural stability--through three complementary tasks. Task 1 (cheque extraction) and Task 2 (Mintaka QA with pinned evidence across nine languages and nine model-effort configurations) test whether accuracy and effort settings transfer; Task 3 (FinanceBench pilot, 50 items over SEC 10-K filings in four languages) validates whether these patterns generalise to financial content. Accuracy reusability is broadly confirmed: Gemini 3.5 Flash achieves 95.2-97.6% extraction accuracy across prompt languages, and FinanceBench per-language means span 5.8pp (77.1-82.9%), though the language-difficulty ranking reverses between domains (Hindi is worst on Mintaka, best on FinanceBench at 82.9%). Effort reusability is domain-dependent: higher reasoning effort yields marginal gains on general-knowledge QA but up to 8pp on financial content, and effort configurations that help in English help proportionally in other languages. Behavioural reusability holds: citation-set overlap with English averages 77-79% Jaccard, with no single language consistently worse. Together, these results support qualified reuse of English-built pipelines, but per-domain, per-language validation remains necessary--neither semantic similarity nor self-reported confidence reliably predicts cross-lingual accuracy.

Candidate-scoring decision models return a probability distribution over a presented set of answers. Public documentation of one such system, Jev, and of an open counterpart, JevAny, describes three training ingredients that determine what these probabilities estimate: soft targets from larger models, a frozen pretrained backbone and temperature calibration. We analyze each ingredient in Gaussian surrogates where the finite-sample risk can be computed exactly. A student trained on teacher targets inherits the teacher's bias, which more student data does not remove and agreement with the teacher does not reveal. A fixed representation lowers the risk only below an explicit sample-size threshold. In a probit model the exact recalibration map is a temperature scaling, which removes all excess risk only when the teacher's coefficient has the correct direction and otherwise leaves a residual that calibration diagnostics do not detect. Under Luce's choice axiom, candidate scores are coherent across refinements of an event if and only if they aggregate by log-sum-exp. On 231,521 credit card complaints from the Consumer Financial Protection Bureau, where the decision is whether a complaint ends in monetary relief, a student trained on a logistic teacher's probabilities ends at the teacher's decision cost while a student trained on outcomes goes 8% below it, a 30B model read zero-shot remains worse than the teacher after recalibration, and that model departs from the coherence condition by 17.5 decibans. Pretrained representations reduce the labeled data required and calibration corrects the scale of the probabilities, but neither corrects a biased teacher, so the probabilities have to be checked against realized outcomes.

Attackers profit from what defenders do not understand. Software security has treated the gap between system behavior and human comprehension as a condition to narrow through analysis and assurance, booking the remainder as accepted risk. Agentic AI systems change that gap: when software reads, writes, explains, and authorizes security-relevant behavior, partial understanding moves faster, grows more opaque, and couples to delegated authority. This paper characterizes the resulting epistemic attack surface and presents a taxonomy of four interacting attack categories, which target the interpretation of intent, the production and consumption of evidence, the scope of delegated authority, and the comprehension of human supervisors. It then defines accountable partial understanding (APU) and specifies six necessary conditions: bounded scope with explicit assumptions, inspectable evidence, assigned responsibility, visible residual unknowns, actionability, and temporal maintenance. It explains why delegating interpretation to software agents differs from distributing knowledge among people and distinguishes APU from risk acceptance, explainable AI, assurance cases, formal verification, and human-in-the-loop oversight. From the conditions, the paper derives five design principles: bounded delegation, decoupled evidence, assumption tracking with invalidation, role-specific accountability views, and intervention-preserving design with residual unknown registers. It introduces two measurable probe dimensions, evidence linkage and invalidation latency, for testing whether accountability mechanisms hold under change and adversarial pressure. Three running examples (an AI coding agent, an autonomous cyber-defense agent, and an AI-mediated insulin delivery update) ground the analysis. Security without full understanding is already here; the task is to make it accountable.

In decentralized federated learning, heterogeneous clients may develop complementary expertise, yet each query is typically evaluated only by the client where it arrives. To exploit this distributed expertise without global coordination, we formulate inference-time collaboration as budgeted local expert referral, a graph-constrained routing process in which a query can move only between neighboring clients and only visited clients contribute to its final prediction. EdgeDRM implements this framework by learning a density-based routing potential for each client: differences between neighboring potentials provide the comparisons used to select successive referrals, while local task models generate predictions that are aggregated along the resulting route. Both training and inference are decentralized, using only neighbor-local information without a central router or graph-wide client scoring. We establish estimation guarantees for the learned comparisons and characterize how local improvement and graph structure govern prior-weighted density gains from referral. Experiments on benchmark and real-world medical data show that EdgeDRM outperforms graph-matched decentralized baselines in most settings, with the clearest benefits under sparse connectivity.

Recent advances in text-to-video generation have enabled high-fidelity single-shot synthesis, yet practical storytelling creation requires multi-shot videos with both strong inter-shot consistency and fine-grained controllability. Existing multi-shot video generation approaches achieve strong consistency but still fall short in controllability. In this paper, we propose Mise-en-Video (MeV), a highly controllable multi-shot video generation framework that allows video creators to conduct staging the multi-shot in the 3D space and then generate the corresponding multi-shot video. To achieve this, our framework combines a hierarchical attention mechanism for multi-level narrative consistency with 3D-aware layout guidance for precise, per-shot spatial control. To support training and evaluation, we develop standardized annotation pipelines and construct a 160K-video dataset with hierarchical multi-shot texts and oriented 3D bounding boxes. Extensive experiments demonstrate that MeV consistently outperforms strong baselines in video quality, inter-shot consistency, and subject-level controllability.

Brain encoding models aim to predict neural responses to stimuli. The most widely used approach fits readouts from frozen pretrained backbones. Using neural data to train large models from scratch or fully fine-tune them after task-pretraining has proven difficult because neural data is costly and limited. We show that parameter-efficient fine-tuning (PEFT) enables us to effectively adapt very large pretrained backbones for brain encoding with a small number of trainable parameters. To make it computationally practical, we exploit the backbone's causal structure to predict multiple timepoints in a single pass instead of using sliding windows, significantly reducing feature-extraction cost. This enables efficient fine-tuning of multimodal backbones with up to 104B parameters, the largest model adapted for brain encoding to our knowledge. Trained with a low-rank adapter (LoRA) and a lightweight readout, our approach consistently improves brain encoding performance across nine diverse multimodal backbones. On the same Qwen3-Omni backbone as the prior state-of-the-art (SoTA), our method BrainLoRA achieves the best published performance on the Algonauts 2025 Challenge. We also show the adapted multimodal backbones remain interpretable: modality attribution recovers the canonical sensory topography that prior work often obtains by construction from separate unimodal encoders, and neural adaptation increases the contribution of audio inputs and cross-modal mixing within the backbone. To scale out the number of adapted models, we develop a multi-LoRA kernel that accelerates joint training of multiple LoRAs on a shared backbone. This enables an ensemble of 200+ adapted models that ranks first on the public Algonauts post-challenge leaderboard at the time of writing. Together, our results highlight PEFT as an efficient and powerful paradigm for brain encoding.

We give a simplified, self-contained proof of OpenAI's FPRAS for counting perfect matchings in general graphs. As in the original proof, we use a restricted Poincaré inequality for two matchings to sample weighted perfect matchings and estimate their total weight. We simplify the proof of this inequality by isolating parts of alternating cycles directly on the original vertex set and controlling the resulting error terms. This avoids the auxiliary vertices and tree constructions used in the original proof. The argument preserves the required parameter bounds and yields the same FPRAS guarantee.

Vision--language models (VLMs) promise to make computational pathology queryable and able to draft reports, but how much language adds beyond the vision backbone is unclear, because almost every system is evaluated by its developers on test sets of their own choosing. We grade published evidence by how it was produced and test four claims with controlled probes that reuse released checkpoints and public data, varying one input at a time. Report text adds prognostic signal: adding a frozen embedding of the stage-redacted report to a slide embedding raises the pooled C-index by 0.019 for progression-free interval and by 0.016 for overall survival across 1{,}633 TCGA patients, and restoring the stage statement adds nothing. Prompts add no measurable accuracy at inference: with features and pooling fixed, zero-shot classification is statistically indistinguishable from a supervised linear probe for lung and renal cancer subtyping. Slide-level question answering is barely grounded: SlideChat answers 74.82\% of SlideBench-VQA(TCGA) items correctly with the slide and 74.41\% without it, whereas its language backbone alone reaches 43.52\%. Paraphrasing leaves accuracy unchanged but flips 8--23\% of answers. Self-reported gains from agentic navigation and few-shot language priors cannot yet be attributed to language. Language thus adds real but narrow value, and we propose minimum controls that would expose these failure modes in future evaluations.

Anonymous peer review protects reviewers from retaliation, yet reviewer anonymity may also be weakened by artifacts released through the review process itself. This risk is growing as many venues require submitting authors to review, creating substantial overlap between authors and reviewers within the same review cycle. We identify review-to-submission linkage, where an anonymous review can be associated with the reviewer's own submission, substantially shrinking the reviewer anonymity set. To assess whether this linkage is practically exploitable, we develop ECHO, which combines technical-expertise and writing-style signals across reviews, papers, and rebuttals, and evaluate it on R2S-Bench, a ground-truth benchmark we construct. ECHO correctly identifies the reviewer's own submission in Top-1 for 34.6% of reviews, in the Top-5 for 54.0%, and in the Top-10 for 63.4%, compared with 0.5%, 2.6%, and 5.3% under random ranking, from the candidate submissions. With a stronger Stage-2 reranker, ECHO further reaches 56.4% Top-1 and 66.4% Top-10. These results demonstrate that review-to-submission linkage is feasible using only routinely released peer-review artifacts. Code and benchmark are available at https://github.com/kaimu11111/ECHO-Revealing-Review-to-Submission-Linkage-in-Anonymous-Peer-Review.

In this paper, we generalize the code basis reduction to non-Hamming metrics. We propose a block basis reduction (BKZ) algorithm for codes that works for all coordinate-additive metrics. Our algorithm is a generalization of the algorithm of Ghentiyala and Stephens-Davidowitz, which in turn is an adaptation of the famous BKZ algorithm for lattices to linear codes and the Hamming metric. For block size~$2$ (LLL), we propose a polynomial-time algorithm with complexity~$\mathcal{O}(n^4 q^3)$ operations in~$\mathbb{F}_q$, where~$n$ is the length of the code and~$q$ is the alphabet size. For the Lee metric we provide an implementation of our algorithms and present experimental results on finding low Lee weight codewords.

Safety alignment for large language models (LLMs) is primarily performed during post-training, while recent work has also begun incorporating safety data into pre-training. Yet safety can degrade when later training no longer includes safety data, and gains in safety often come at the cost of general capability. This tension suggests that safety and general reasoning should not be treated as separate training goals. We instead view safety as an LLM capability that can be learned through a reasoning structure shared with general tasks. Both require the model to understand intent or task goals, apply relevant knowledge, and reason about potential consequences. We introduce safety mid-training, a stage between pre-training and post-training that jointly develops safety and general reasoning through this shared structure. Given safety and utility prompts, the base model generates structured documents and reasoning traces that are then used for mid-training. Safety mid-training reduces average attack success rate (ASR) by 72% to 81% relative to Qwen3-1.7B, 4B, and 8B base models, while also improving general capability. The benefits persist through subsequent post-training. After general-only SFT, mid-trained models largely retain their safety, achieving lower ASR than the baselines without sacrificing general capability. They also retain safety under general-only GRPO without safety rewards, while subsequent RL reaches a better safety-utility frontier when initialized from the mid-trained models. Together, these results suggest that learning safety and general reasoning within a shared mid-training stage can reduce the safety-utility trade-off and make safety more robust to subsequent training.

This paper presents the hardware implementation of a quadratic program-based controller on Harpy: thruster-assisted bipedal robot to perform thruster-assisted trotting. This is inspired by wing-assisted locomotion observed in birds such as the chukar partridge, which employ thrust vectoring to manipulate the ground reaction force. The proposed controller formulates the thruster-assisted spring-loaded inverted pendulum dynamics as a condensed QP-based model predictive control problem, enabling simultaneous regulation of ground contact force and active thrust vectoring at 500 Hz. A decoupled control architecture is used, where the robot's sagittal dynamics are controlled using QP-MPC while the frontal dynamics is regulated using a cascade orientation stabilization controller. Two experiments were conducted with different forward reference velocities of 0 m/s and 0.1 m/s. While the instantaneous velocity oscillates, the mean velocities can track their respective reference. Additionally, joint tracking error remains within 1 deg while tracking the desired leg length given by the leg trajectory controller and desired swing trajectory given by Raibert's heuristic. The QP-MPC problem is consistently solved with 0.18ms, which is well within the 2ms computation budget while satisfying the GRF friction cone constraint.