Science & Mathematics

Safety risks from fine-tuning are commonly characterized by the harmfulness of the training data:harmful fine-tuning is considered dangerous, while benign fine-tuning is often assumed to preservealignment. We argue that this view misses a more fundamental factor-the behavioral transition inducedby supervision. Harmful fine-tuning and safety-degrading benign fine-tuning can both be understoodas learning the same refusal-to-compliance transition, but on different data distributions. We studywhether such a transition learned on benign inputs can generalize to harmful requests. We constructbenign examples that an aligned model initially tends to refuse and pair them with their originalharmless responses, creating a discrepancy between the model's initial refusal and the supervisedanswer. Fine-tuning on these examples increases harmful compliance, but only when two conditionshold together: the discrepancy is large enough, and it cannot be resolved through a narrow surface-levelshortcut. With the induced discrepancy matched, varying only whether it is localized changes harmfulcompliance by an order of magnitude; raising the discrepancy does not help while a shortcut remainsavailable, and removing the shortcut does not help while the discrepancy is small. Building on thisobservation, we introduce SAFEOFF, a dataset of benign instructions rewritten to raise the alignedmodel's refusal likelihood and paired with their original answers. Across three aligned LLMs and threesafety benchmarks, SAFEOFF substantially weakens safety alignment despite containing no explicitharmful supervision, suggesting that fine-tuning safety depends not only on what data a model sees, butalso on which behavioral boundary the supervision moves.

State space models (SSMs) combine fast inference and fixed state size with competitive language-modeling performance. Recently, state souping has been proposed as a procedure whereby an SSM independently encodes chunks of tokens like documents, tool definitions, or memory records. The resulting state encodings may be cached and recombined to cheaply construct a context for downstream queries. However, both state souping and vanilla SSMs deteriorate on long, crowded contexts, limiting their utility for long-horizon domains. To address this issue, we propose weighted state souping, a simple method that learns query-dependent weights for composing cached states. The resulting models improve long-context performance substantially across both single- and multi-hop document question answering, needle-in-a-haystack retrieval, variable tracking, and tool use. Our method can be further adapted into an efficient sparse variant, yielding near-constant encoding cost with respect to context length while preserving (and sometimes improving) performance. We also present simple theoretical arguments that illustrate why souping works and when it may fail, and discuss informally the advantages of weighted souping. Weighted state souping thus combines reusable context representations with selective retrieval, offering an efficient approach to language modeling over large collections of documents, tools, and other forms of structured context that increasingly dominate modern agentic workloads.

Vision-based object pose estimation is vital for conducting spacecraft rendezvous and proximity operations (RPO). Recent research has demonstrated the robustness of event cameras against the harsh lighting conditions that often occur in orbit, thanks to the extremely high dynamic range of event sensors. However, event-based spacecraft pose estimation (EBSPE) suffers when the relative motion between the chaser and target is slow, since the event rate drops significantly, leading to data with low signal-to-noise ratio. Unfortunately, for safety reasons, real RPO trajectories often involve slow relative motions (e.g., down to 1 mm/s), and operating under low illumination exacerbates the problem. Addressing slow relative motions is thus crucial to raise the practicality of EBSPE. To this end, we propose using recurrent vision transformers (RVT) for EBSPE. The key idea is to exploit recurrent memory to capture longer-term information across sparse event streams, and to learn the temporal attention span from full trajectories rather than fixing the time window. We developed an RVT-based heatmap regressor to produce 2D-3D correspondences for pose estimation, which was validated on real event data collected under slow relative motions representative of RPO trajectories. Results indicate the superior pose estimation accuracy and stability of our RVT-based method, relative to previous per-event-frame-based or fixed-time-window EBSPE methods.

Cheng and Wang (CCC' 24) showed that BPL is contained in logspace uniform AC$^1$ (denoted as L-AC$^1$). We upgrade their result by showing that approximations of $\#$SAC$^1$ functions can be computed in L-AC$^1$ with multiplicative error $2^{-\mathrm{poly}\log n}$. Previously, approximations of $\#$SAC$^1$ functions are known to be in $\mathsf{NC}^2$ via its exact computation given by Vinay (CCC' 91). A main corollary of our result is that Probabilisitic polynomial time logspace AuxPDMs and Read-$k$ logspace PTMs are computable by L-AC$^1$ with multiplicative error $2^{-\mathrm{poly}\log n}$. Notice that multiplicative error approximations are stronger than additive error approximations in the sense that they can be used to decide both non-deterministic computations and two-sided error probabilistic computations. We note that recently OpenAI announced that BPL $=$ L. Even if we only consider derandomizing probabilistic logspace, our L-AC$^1$ computation has the advantage that the error can be $2^{-\mathrm{poly}\log n}$ and can be multiplicative.

Reinforcement learning (RL) has emerged as a key method for training AI models in scientific applications, but RL remains difficult when outputs cannot easily be verified against a known ground-truth answer. Natural-language weather forecast reports are one such domain: they translate atmospheric data into assessments of weather evolution, hazards, and uncertainty for both the general public and key decision-makers. Unlike numerical weather-analysis questions, they admit many valid outputs, which limits deterministic verification. To address this, we develop Zephyrus-R1, a practical framework for RL on scientific weather forecast reporting. We focus on three key components of the agentic RL pipeline: reward design, curriculum learning, and an analysis of human-expert alignment. To encourage both report legibility and factual accuracy, we design a reward function that combines a meteorological rubric with bidirectional claim entailment, both evaluated against National Weather Service reference reports. Our framework optimizes this reward in a simulated tool-calling environment, enabling vision-language agents to gather and analyze forecast evidence across regions and lead times before producing a report. We use annotations from three meteorologists to audit the claim labels produced by our LLM judges. Finally, a curriculum helps small models first develop tool use and scientific reasoning through RL on verifiable weather-analysis questions, then optimize the open-ended report reward. After training Qwen3.5-4B with Zephyrus-R1, we improve the mean across six judge-metric scores by 26% over the base model and achieve results competitive with frontier open-weight models. For Qwen3.8-27B, we improve the same mean by 13% over the base model, by 6.7% over Kimi-K3, and by 2.7% over GPT-5.6-Luna.

Knowledge acquisition from black-box large language models (LLMs) offers a practical approach to adapting smaller student models. However, many existing methods rely on predefined queries or student feedback to guide acquisition, leaving autonomous expansion of knowledge targets an important challenge. We propose \textbf{MapSteal}, a student-independent framework for structured recursive knowledge acquisition. Starting from a small set of initial roots, MapSteal represents acquisition targets using structured states and recursively expands them through teacher-proposed directions. Question realization, proposal admission, and deduplication constrain the acquisition process, while root-aware scheduling organizes exploration under a fixed QA budget. The resulting question--answer pairs are used to adapt student models without requiring student feedback during acquisition. Experiments with Qwen3.5 students at 2B, 4B, and 9B scales show competitive downstream performance across six medical and cybersecurity benchmarks. On the 4B medical evaluation, Full MapSteal improves average accuracy by 1.17 percentage points over a simplified recursive acquisition variant. Cross-domain evaluations show that gains depend on the domain and evaluation panel. These results demonstrate the feasibility of structured, student-independent recursive acquisition while highlighting the distinction between generating acquisition targets and obtaining useful training supervision. Our code is available at \href{https://github.com/LabRAI/Mapsteal} {https://github.com/LabRAI/Mapsteal}.

Personalized public-health targeting must capture heterogeneous intervention benefits while remaining interpretable under resource constraints. Flexible causal-effect models can be hard to operationalize, while coarse subgroup policies may lose individualized effect information. We study budget-constrained hypothetical state interventions in observational data without a logged historical decision policy. For obesity, elevated glucose, and lifetime smoking-history states, causal-structure-informed pretreatment covariates are prespecified, with capacity capped at 70% of those observed in the adverse state. We evaluate causal distillation where Causal Forest and Bayesian Causal Forest teachers estimate conditional average treatment effects (CATEs), which decision tree and random forest students approximate. We compare individualized policies with unsupervised K-means subgroup policies and supervised causal-tree policies with leaf counts matched to K-means cluster counts. Observational policies are assessed on held-out data using a doubly robust estimator, bootstrap uncertainty, and decision-focused teacher-student fidelity measures. Across PIMA obesity and glucose and NHANES smoking-history experiments, each distilled student was within 0.011 absolute utility of its teacher, with allocation Jaccard similarities of 0.738-0.937. K-means and causal-tree policies were also competitive, with the highest point-estimate policy varying by application. Prespecified paired comparisons were nonsignificant after Holm adjustment, while near-budget-boundary agreement was less stable than global fidelity for some teacher-student pairs. Findings support causal distillation for translating flexible effect models into simpler budgeted targeting rules, although observational estimates depend on causal-identification assumptions and require longitudinal or prospective validation.

Humanoid robots are envisioned to perform diverse human-level tasks through their anthropomorphic morphology. Their dual-arm configuration enables a broad range of everyday actions. Yet, as in humans, this morphology can constrain loco-manipulation when both arms are occupied or additional contacts are required beyond their reach. Supernumerary robotic limbs (SRLs) offer a way to extend these capabilities, but their integration requires the added limb to be coordinated with whole-body motion and balance. In this study, we develop and integrate a supernumerary arm into a humanoid platform to investigate how morphological augmentation reshapes its feasible action space. Through quantitative analysis and hardware experiments across two configurations, we show how the additional limb expands whole-body interaction capabilities by enabling manipulation and support to be coordinated beyond the constraints of the native dual-arm configuration. To control this augmented morphology, we leverage model-based trajectory optimization (TO) to generate dynamically feasible motions for goal-conditioned reinforcement learning (RL) controllers. The controllers are trained in simulation and transferred to hardware in a zero-shot manner, enabling coordinated whole-body skills for contact-rich tri-arm loco-manipulation. We further show vision-guided autonomous execution on selected tasks using diffusion policies trained from hardware demonstrations. Our findings suggest that humanoid morphology need not be restricted to human-like forms, but can be reimagined as an extensible design space for robotic augmentation.

Agent pattern catalogs describe architectural alternatives, but engineers also need evidence for selecting and composing them. We extend a cognitive-function--execution-topology framework with an evidence-driven selection procedure and explicit composition contracts. The study combines two industrial case accounts, a contributor's collaboration design, seven practitioner workshops, and executable fault-injection experiments. Three local reference systems use SQLite state, shared artifact files, and a file-backed map-publication surrogate. A 2 x 2 ablation separates function-specific contracts from execution controls across 36 scenario families, with ten input variants per family. Under equal family weighting, composition correctly completes or safely refuses 11 of 12 families in each system, compared with 2, 2, and 1 for the simple baselines. These are constructed challenge-suite results, not production success rates. Recovery without consumer-side verification increases false-success cases in the map surrogate. All composed systems still fail a boundary case in which trusted evidence is wrong. An audit of a frozen six-arm language-model pilot reproduces its scores but confirms an evaluator-dependent ranking reversal. The contribution connects industrial failure, pattern selection, executable treatment, external acceptance, and revision. Selection is demonstrated rather than evaluated against an alternative procedure; catalog admission rests on separate criteria. The results support bounded composition mechanisms, not a claim that the framework improves engineers' design decisions over an equally informative conventional method.

The macroscopic capabilities of intelligent systems arise from the local interactions of many simple units. When such systems are deployed in the real world, their behaviour must nevertheless be constrained, leaving designers caught between releasing adaptability and preventing loss of control, yet without formal tools to resolve the dilemma. This paper proposes a unified research program for controlled emergence that restates this tension as a pair of dual problems: the forward problem explains macroscopic behaviour from local rules, and the inverse problem solves for controllable rules and readout mechanisms from a desired target. We take an augmented quintuple as the formal object shared by both ends, and decompose ``engineering'' into three independently measurable coordinates: controllability, reproducibility, and degree of formalizability. On this basis we give three testable working hypotheses (critical scale; the cooperation and trade-off between constraint and adaptability; minimal constraint) and one feasibility criterion, which reduce inverse design from unconstrained optimization to interval checking. The program takes explicit failure boundaries as its premise, does not pursue complete control, and always preserves the system's room for self-organization. The reduced pilot experiment attached to this paper serves only to verify the executability of the research protocol and does not constitute a statistical test of the working hypotheses. This design language generalizes to multi-agent artificial intelligence, the industrial internet, and emergency response systems, all of which must confront emergence and controllability at once.

Camera-based 3D semantic occupancy models can be trained without 3D labels by rendering their predictions into the images and supervising them with 2D pseudo labels. In sparse Gaussian representations trained this way, Gaussian positions are learned only through pixel-wise rendering losses, which we hypothesize give small foreground objects little weight. In trained GaussianFlowOcc models, the Gaussian density on foreground voxels is on average only 1.62 times that on background voxels. We propose a training-only foreground attraction loss that back-projects foreground pseudo-label pixels with pseudo depth and pulls nearby Gaussians toward these 3D points, favoring Gaussians that already predict the target class; the inference network is unchanged. On Occ3D-nuScenes, the loss raises this concentration to 2.40 over three seeds. Over three seeds, the mean foreground IoU increases from 6.29 to 6.96 and is higher for each seed, whereas the change in mean mIoU (13.53 to 13.82) is within the seed-to-seed variation of the baseline. Single-run ablations further indicate that pseudo-label targets outperform targets placed at ground-truth voxels and that acting on the Gaussian positions alone is not sufficient, and a single-seed test shows that the model degrades less when fewer Gaussians are kept at inference.

Improving pretrained robot action models at test time from only a few rollouts remains challenging. Activation steering offers a lightweight approach by modifying feature activations while keeping model weights fixed. However, effective steering requires identifying which features to modify and determining the timing and strength of modification. We introduce SteerAct, a few-shot temporal steering framework that uses a shared temporal descriptor of sparse-feature dynamics for offline feature selection and online control. A pretrained sparse autoencoder provides candidate features, and our temporal descriptor summarizes the distribution of local activation-order patterns for each feature, enabling comparisons across rollouts without requiring timestep alignment. Before steering begins (offline), we use successful and failed reference rollouts to select a single task-specific feature whose temporal descriptors best distinguish the two outcomes and construct its reference banks. During each new execution (online), we compare the feature's evolving temporal descriptor with these references to determine when to intervene and how strongly to increase or decrease its activation via a steering coefficient. Using only 24 outcome-labeled rollouts per task, SteerAct improves task success by approximately 10% relative to unsteered models in evaluations spanning $π_{0.5}$, Fast-WAM, RoboTwin 2.0, DexJoCo, and real-world manipulation, without updating action-model weights. Code and demos are available at: https://steeract.github.io/.

Unsupervised cross-domain person re-identification remains challenging because domain shifts generate noisy pseudo labels that are reinforced during iterative adaptation. Existing clustering-based methods often rely on limited reliability cues or global confidence criteria and may impose hard identity supervision on ambiguous target samples. To address these limitations, we propose CGPLR, a confidence-guided pseudo-label refinement framework for unsupervised domain-adaptive person re-identification. CGPLR estimates sample-level reliability by jointly modeling centroid compatibility, local neighborhood consistency, separation from the strongest competing cluster and temporal assignment stability. The resulting confidence distribution is used to derive cluster-adaptive upper and lower thresholds for a three-way selective decision. Reliable assignments are preserved, recoverable samples are reassigned only when an alternative cluster receives sufficient geometric and neighborhood support and highly ambiguous samples are deferred from hard identity supervision. The learning objective adapts to these decisions through confidence-weighted supervision for preserved samples, cautious corrected-label weighting for reassigned samples and label-free representation consistency for deferred samples. CGPLR achieves 76.4% Rank-1 accuracy and 47.6% mean average precision (mAP) on Market-1501 to MSMT17, and 95.1% Rank-1 accuracy and 87.7% mAP on MSMT17 to Market-1501, demonstrating effective reliability-aware selective adaptation. The code is available at https://github.com/Sundas058/CGPLR.

Autoregressive (AR) video diffusion models generate streaming video of arbitrary length by conditioning each new chunk of frames on those already generated. However, they suffer from exposure bias, where a model trained on clean context must continue from its own imperfect outputs, and errors compound until quality collapses within seconds (drift). Existing methods either require cluster-scale base retraining, apply training-free inference corrections with inconsistent results, or train frozen-base correctors for sampling and cache errors. We instead find that drift is largely deterministic and therefore learnable, as holding the noisy state fixed while varying only the history reveals a counterfactual velocity gap that is ~95% systematic across noise seeds. Based on this observation, we introduce Clean Forcing, which trains a 5.9M-parameter LoRA corrector on the frozen base model to regress this gap, with a closed-loop stage that adds DAgger-style aggregation and a drift-contraction objective to expose the corrector to its own rollouts. Because the clean histories can be the base model's own single-shot generations, training needs no real videos, and the merged corrector adds no inference cost. In 50 s text-to-video generation, Clean Forcing reduces Δ-drift from +11.3 to +1.65 without real video and to -0.16 with 40 real clips, improving on the best published result of +3.57 with over 150x less data. Extensive experiments demonstrate that Clean Forcing outperforms prior methods in aesthetic quality and data efficiency with reduced drift for long-horizon video generation.

Structural MRI-based classification of major depressive disorder (MDD) remains challenging due to subtle, heterogeneous anatomical changes. We propose uncertainty-gated SHAP-Dirichlet refinement, a post-hoc reliability layer that improves predictions without retraining the classifier. The method retrieves similar labeled SHAP explanations, constructs Dirichlet class evidence, and selectively refines uncertain predictions through probabilistic fusion, while allowing unreliable cases to be deferred for expert review. We evaluate five classifiers on 504 SRPBS subjects (252 controls and 252 MDD patients) using 159 anatomical features and repeated nested cross-validation. Refinement improved accuracy across all 8 repetitions for TabM and ResNet, yielding mean gains of 1.04 and 0.45 percentage points (sign test p=0.008), while modifying at most 8.3% of predictions. Gains were concentrated among low-confidence cases, with significant improvements in calibration. The method also benefited an RBF support vector machine but not logistic regression or random forests. Ablations confirmed the importance of uncertainty gating and soft fusion. Dirichlet evidence identified baseline errors (AUROC 0.62), while deferring 50% of cases increased retained-case accuracy to approximately 75%. These findings support selective explanation-guided refinement as a practical approach to improving prediction reliability and facilitating expert triage in MDD classification.