Papers

Video diffusion models generate visually compelling short-term sequences but often struggle to preserve scene geometry and object structure over time. In this paper, we first demonstrate that current video attention cannot identify reliable correspondences across frames, with responses dispersed across multiple regions even after recognizable scene structure has emerged. We argue that reliable cross-frame associations are essential for information exchange as scene structure forms in video generation. Thus, we introduce Geometry Attention Distillation (GAD), which transfers correspondences from a frozen geometry model's attention to video attention. Extracted from clean videos, these correspondences link the same static surface or moving object across frames. Since structure formation begins in deeper layers at higher noise and shallower layers at lower noise, we apply attention distillation at a student layer selected according to the noise timestep. After training, video attention focuses more consistently on corresponding locations. Experiments across all three video backbones demonstrate improved long-range scene consistency and dynamic temporal consistency, with geometric gains across generation lengths. On Wan2.2 5B, GAD improves cross-frame matching accuracy by at least 30% over the Wan baseline on both long-range benchmarks. Notably, GAD achieves these gains without additional inference modules, offering a promising approach to efficient and consistent video-based world simulation. Project page: https://fantasy-amap.github.io/GAD/.

We study the randomised and deterministic demand-query complexity of approximate bipartite matching with few query rounds. The demand-query model, introduced in [Blumrosen and Nisan SICOMP 2009; Nisan SODA 2021], is inspired by combinatorial auctions and welfare maximisation, where an oracle returns the first neighbouring vertex (item) of a vertex u (agent) according to a given permutation pi (preference ordering of the agent over the items). For randomised algorithms, our main result gives a trade-off between the number of rounds k, the per-round query budget q=n^{1+\delta}, and the achievable approximation ratio. Let 1 <= k <= log n, and -1 <= delta <= 1/k. Every k-round randomised algorithm using O(q) demand queries per round has a worst-case expected-size approximation factor of Omega(n^{1/2k-delta/2}/k). Conversely, we give a randomised algorithm achieving approximation factor O(n^{(1-k*delta)/(k+1)}/k) in expectation. For deterministic algorithms, for k*q >= n, we show that the optimal deterministic approximation factor is ~Theta(n^2/(q*k^2)), thereby improving the previous lower bound of [Nisan SODA 2021] by a polynomial factor over a wide range of parameters. In particular, for q=Theta(n), we improve the deterministic round-complexity lower bound for constant-factor approximation from ~Omega(n^{1/3}) to ~Omega(sqrt n). Our algorithms produce matching and vertex-cover certificates; the deterministic trade-off also extends to minimum vertex cover and maximum fractional matching. We extend the randomised upper bounds to the weaker Demand-OR model [Nisan 2021]. We prove a non-adaptive simulation of uniform random-neighbour queries using known techniques, avoiding the O(log n)-round penalty of the trivial binary-search simulation. Consequently, our upper bounds yield the same k-round approximation guarantees using O(q * log^2 n) Demand-OR queries per round.

A robot needs a world model that predicts how its actions will change the scene and a policy that chooses those actions. Models are usually built and evaluated for only one of these roles, and every robot records its actions in its own format. We ask whether one robot world model can serve both roles across different bodies, and what makes this work. We present GreenWAM, a world foundation model for robotics, together with the recipe that produces it: (1) a shared anatomical action space that describes 20 robot families, from bimanual arms to dexterous humanoids, in common joint coordinates; (2) joint training on these families in visual, action-conditioned and world-action modes; and (3) policies that split an action expert off the generator. As a world model, GreenWAM attains top-tier trajectory, depth and 3D accuracy on WorldArena 1.0, ahead of every world-action model and video generator in our comparison. As a policy, it scores among the top entries of the RoboTwin 2.0 leaderboard, reaches the highest zero-shot LIBERO-plus success among the compared methods and is competitive with dedicated vision-language-action and world-action policies on CALVIN and LIBERO. We further adapt GreenWAM to a real humanoid, where it succeeds more often than policies adapted from $\pi_{0.5}$ and from Cosmos 3 Edge, and distill it into a single-step policy for real-time control. Controlled experiments yield three findings: robot specialization improves world prediction and control together; an action expert helps when split off at the start of adaptation; and keeping world-model modes during policy adaptation preserves robustness to scene changes.

System one models enable direct, low-latency decision-making. Yet recent practice of predicting an answer from serialized text does not, by itself, respect the structure of a decision problem. We argue that current designs underemphasize four important structural requirements. (i) null-aware: Models should distinguish answer absence from uncertainty among candidates; (ii) permutation-invariant: Reordering alternatives should not change the decision; (iii) option-scalable: Growing candidate sets should not require redefining the output space; (iv) cacheable: Request-independent computation over recurring alternatives should be reusable across requests. We introduce Jeva, a decision-native architecture that addresses these requirements jointly. A shared encoder independently represents the request and each option, while a compact decision stack refines only the request state against immutable option memory. Symmetric aggregation yields permutation-equivariant probabilities; shared scoring accommodates variable-size candidate sets. A separate coverage estimate distinguishes answer absence from candidate competition, enabling explicit null outcomes. Request-independent option representations and layer projections support reuse across newly assembled sets. These mechanisms unify Choice, Noul, and Score outputs. Across evaluations, Jeva combines strong typed-decision performance with near-perfect null detection and no decision flips under any tested permutations. It scales to thousands of candidates and substantially reduces latency through cross-request reuse without changing decisions.

Limited battery endurance remains a major obstacle to persistent UAV missions. Existing energy-replenishment approaches typically rely on battery-swapping/charging infrastructure or on ground vehicles with accessible road networks - assumptions that may not hold in infrastructure-less regions (ILRs) such as remote wilderness, mountainous terrain, and disaster-stricken areas. We formulate energy resupply in ILRs as a distinct logistics problem, in which energy itself must be projected into a region with no pre-existing infrastructure, and proposes a Distributed Energy Relay Framework that enables cooperative energy delivery using porter drones. The framework introduces a Transportation-by-Relay-and-Cache (TRC) mechanism, in which drones relay battery resources toward a mission location. Additional models are developed to characterize battery transportation capacity, relay amplification, energy-delivery capability, operational cost, and reliability. Based on these models, stations independently evaluate their delivery capability and submit bids to a mission coordinator, which selects an energy-delivery coalition that satisfies mission requirements while minimizing cost. Analytical evaluation illustrates the impact of battery characteristics and drone capabilities on energy projection and demonstrates the effectiveness of the proposed framework for sustaining UAV missions in ILRs.

International challenges have become the standard for developing and benchmarking biomedical image analysis methods using centralized multi-institutional data. Federated learning (FL) describes a paradigm-shift on multi-institutional collaborations, where models are trained across institutions while their data are always retained locally. The Federated Tumor Segmentation (FeTS) challenge, first held in 2021, was the first challenge dedicated to FL in medical imaging and editions in 2022 and 2024 pushed the frontiers of real-world method validation. FeTS++ 2027 returns after a two-year gap with a substantially larger and more diverse cohort, and with the two "pluses" that name the challenge: an official classification task alongside segmentation, and a move of the infrastructure to NVIDIA FLARE (NVFlare). Built on the clinically acquired, multi-institutional multi-parametric MRI data of the BraTS 2025 Lighthouse challenge, FeTS++ spans gliomas and meningiomas across patient cohorts from institutions in North America, Europe, Asia, and Africa, and focuses on ranking submitted methodological approaches according to their model utility and an efficiency-aware score. Task 1 asks participants to find the best strategy for aggregating knowledge from per-institution segmentation models, while Task 2 asks them to distinguish glioma from meningioma under the same federated protocol, in which the entity label is heavily confounded with the site of origin.

Predicting spatial gene expression from standard H&E images could extend molecular analysis to tissues without spatial transcriptomic measurements. A standard Transformer applies self-attention over spots but ignores where each spot sits in the tissue, whereas a spot's molecular state depends on tissue context at multiple spatial scales. Building on this insight, we introduce MIST, a Multi-context Spatial Transformer that augments the self-attention backbone with two context pathways applied in every layer: a coordinate-invariant local pathway that attends over spatial k-nearest-neighbor spots, and a slide pathway that broadcasts a whole-slide summary to every spot; the model predicts all spots in a single forward pass. Because coordinates enter only through pairwise distances, the local pathway is invariant to translations, rotations, and reflections when image features are fixed. In intra-cohort evaluation across six human organ cohorts from HEST-1k, MIST improves mean Pearson correlation over the strongest evaluated baseline by 9.7%, 11.6%, and 11.4% for 10-, 50-, and 100-gene panels, respectively, with the lowest average MSE and MAE. MIST further improves leave-one-organ-out transfer by 6.0% over the next strongest baseline, attains the best expression-derived domain recovery (highest mean ARI and NMI at every clustering resolution, and higher NMI than the closest baseline on 84% of held-out slides), and reproduces spatial expression patterns of selected cancer-marker genes. Together, these results show that adding local and slide context to a vanilla Transformer backbone is an effective approach to spatial gene expression prediction from histology.

In this paper, we study online paging with delay or deadlines. In online paging with delay, an online algorithm manages a cache of size $k$ and needs to serve incoming requests by adding the requested pages to the cache. In addition, it has the option to delay serving incoming requests at a cost that depends on the delay. We propose a new online algorithm, which combines a threshold rule with first-in, first-out (FIFO) replacement, and show that it is $(2k+2)$-competitive for cache size $k$. This result corrects a previously published claim that the competitive ratio is exactly $2k+1$ regardless of the number of holes (pages outside the cache). However, in the case of one hole, we show that the FIFO rule with a different threshold is indeed $(2k+1)$-competitive, giving a tight result. In the case of deadlines, the same analysis gives a deadline-triggered algorithm that is $(k+1)$-competitive in general and $k$-competitive if there is only one hole. Moreover, we show that the second hole makes paging with deadlines strictly harder than classical paging, where the competitive ratio is independent of the number of holes. For paging with deadlines, we prove a deterministic lower bound of $k+\tfrac12$, already for two holes. We show that, as a consequence, no threshold-triggered algorithm can achieve a competitive ratio less than $2k+\frac32$ in general.

Rubric-based reward models (RMs) evaluate responses against explicit criteria. However, existing methods often derive criteria from observed response candidates, introducing response-induced bias that may limit rubric generalizability. Their judges also struggle with mechanical constraints when relying on reading alone, and commonly used pairwise judges provide only relative preferences, complicating their direct use for policy optimization. We propose ReVeR, a pointwise rubric RM that derives hard rules and graded principles by reasoning about the instruction alone and evaluates responses independently through tool-augmented judging. Both components are trained through label-blind distillation, in which preference labels only filter a teacher's outputs, yielding ReVeR-Traces, a large-scale dataset of rubric-reasoning and judging trajectories. Experiments demonstrate that ReVeR outperforms pairwise rubric RMs in average preference accuracy, most clearly on instruction following, as well as in test-time scaling and downstream policy optimization. We further find that generator RL improves preference accuracy while reducing rubric fidelity, and that higher preference accuracy does not necessarily translate into better downstream policies. These findings highlight that rubric fidelity, preference accuracy, and downstream utility are distinct properties, shedding light on future RM research.

Human-AI Interaction is commonly evaluated by what an artificial agent enables during use, yet interaction can also alter the human who acts next. We introduce Residual Human States, a trajectory-based framework for representing the potentially multidimensional human state carried forward after interaction with AI, relative to a relevant alternative trajectory. The framework distinguishes assisted output from human-state transition and derives four empirical propositions: multidimensionality, interaction-policy dependence, horizon dissociation, and temporal persistence. Its central claim is that assisted performance does not by itself determine the human state carried forward. We examine these propositions through four empirical cases with different inferential roles. Published medical-training evidence illustrates how one AI interaction history can be associated with both slightly higher observed accuracy on some later unaided cases and transmission of a systematic decision bias on others. A second diagnostic-learning study shows that later unaided behaviour differs across AI-support policies, while also revealing that aggregate accuracy cannot by itself distinguish improved discrimination from a shifted response criterion. In a classroom experiment, generative-AI tutoring substantially improved assisted practice performance without a corresponding detectable gain on a subsequent unaided exam. A longitudinal reanalysis of personalised AI dialogue further shows that residual contrasts can evolve across post-interaction horizons and that such change may arise from the comparison trajectory as well as from the focal intervention. Evaluating interactive AI systems therefore requires attention not only to contemporaneous performance, but also to which human states alternative interaction trajectories produce, relative to which counterfactuals, and how durably.

Ghostline is a device-free motion tracking system built on six ESP32-S3 Wi-Fi boards and one consumer router. The boards' Wi-Fi driver reports channel state information (CSI) for every received frame with the sender's address, and a modified open-source firmware forwards both, so every board carries a CSI stream for the router and for each other board: seven radios give up to 36 directed Links over 21 Spans instead of a star of six. Each Link is compared with a self-forming quiet Baseline, disturbed Links vote on a floor grid, and the grid becomes anonymous Tracks. The system was built and run during a three-day hackathon in an ordinary room. We describe it in full and report a pilot evaluation with one walker: spoken cues at seven marked positions, a two-minute free walk and five minutes of empty room, synchronised to video within 0.18 s (median). Walking between marks, the Track lay a median 0.85 m from the straight path and 1.56 m (95% CI 1.29 to 2.83) from a cue-based reference trajectory (constant speed between cues, not measured positions), with 46% of walk updates within 2 m of it; the free walk was tracked in 94% of updates. The limits: a standing person is lost within seconds, a false Track appeared in 15.9% and 2.9% of updates in two empty-room periods, and informal runs that afternoon did not separate two walkers. Two offline replays through the same pipeline show where this comes from: with the node-to-node Links removed and settings unchanged, the tracked share of walk updates falls from 90% to 19% and the free walk from 96% to 60%; a plain radio tomographic image on the same per-Span evidence, at the same false rate, matches the grid and tracker at 2 m (48% against 47%) and is ahead at 1 m (36% against 21%). On this recording the Tracks depend on the Link set, and the imaging stage adds no measurable accuracy. We state what the pilot establishes and what it does not.

Online learning platforms are increasingly deploying AI assistants to provide personalized, on-demand tutoring. These can leverage infographics to communicate complex information and improve the learning experience. However, automating effective infographic generation remains a challenging task, requiring a combination of visual richness with precise typography, structured layouts, and continued editability. Existing representation methods satisfy these goals only in part: Scalable Vector Graphics (SVG) representations preserve geometry, hierarchy, and text but provide limited aesthetic flexibility, while raster diffusion models produce richer appearances by flattening content and layout into a single image. Conditioning a diffusion model on a rasterized SVG does not resolve this conflict because the raster preserves approximate placement but discards element ownership, occlusion relationships, and the structure needed to modify and recompose the result. We introduce LayerCanvas, a hybrid approach that retains SVG as the semantic, structural, and typographic program while using diffusion to aestheticize non-text content. Rather than generating one flattened raster, LayerCanvas partitions SVG elements into z-ordered sets of mutually non-overlapping elements and processes them through communicating parallel RGB generation streams. The resulting layers are independently editable and deterministically recomposed, while text is rendered from preserved vector metadata. Our approach combines rich visual synthesis with accurate typography and localized editing. LayerCanvas substantially improves text fidelity over state-of-the-art raster image generators while achieving the highest overall aesthetic quality. We further demonstrate non-destructive content editing, localization, asset reuse, and animation.

Evaluation for automatic crisis reporting should measure what practitioners need, not what is convenient to compare. Through LLM-assisted, author-approved content analysis, we derive a 15-slot information schema from professional humanitarian situation reports using qualitative content analysis, then evaluate 121 human and 261 LLM-generated reports across five disasters and three hazard types. Under google/gemini-3.7-flash on Cyclone Idai, the best machine arm is nominally above the empirical human ceiling (0.827 vs. 0.821), but the difference is statistically indistinguishable from zero (95% bootstrap CI $[-.049,.059]$; permutation $p=.852$), and the ordering does not replicate under openai/gpt-5.6-luna. On a blind 100-item Idai sample, the author agrees with the Gemini judge at 0.910 accuracy and unweighted Cohen's $\kappa=.799$. All four judges clear the plan's raw-agreement criterion of .7; only Gemini reaches .7 on chance-corrected $\kappa$. Across the three primary events, machine reports reproduce 5.4-11.0% of same-day official figures. Across those same events, raw exact-number matching links 14.3-19.6% of official figures and 85.4-90.1% of machine figures to the social stream; manual audit confirms 45/50 declared absences, 4/25 official matches, and 24/25 machine matches semantically. The issuing-provenance gap is positive in all 12 observed event-judge cells: report date, organisation, series, and authority were not supplied at professional-report coverage in any cell. On the three primary events, ROUGE-L is uninformative, BERTScore is negatively associated with completeness after multiplicity correction, and a generic LLM judge changes sign.

Test-time scaling for large language models (LLMs) can involve generating multiple candidate solutions for selection or revising an earlier attempt. To guide selection and revision, we need a verifier that can estimate candidate correctness and provide evidence of failure. In competitive programming, however, official judging systems provide authoritative verdicts under finite submission budgets, without revealing failure explanations or failing inputs. To provide informative feedback without consuming official submissions, we formulate a task in which a model judges and diagnoses a candidate program using only the problem statement and source code, with generated counterexamples serving as checkable evidence. To evaluate this task, we build USACO-JUDGE from the USACO 2026 contests, using accepted and unaccepted programs written by human experts and multiple LLMs. We then train ARBITER, a 35B mixture-of-experts (MoE) agentic verifier, with reinforcement learning in a generic execution sandbox. Training raises counterexample success on incorrect programs from 40.61% to 62.67% and mean binary verdict accuracy on USACO-JUDGE from 71.41% to 87.41%, approaching frontier verification performance. Beyond verification, ARBITER outperforms baselines in both code selection and refinement. In an IOI 2026 case study, ARBITER-guided refinement raises GPT-OSS-120B's expected score above the gold-medal threshold under the official submission limit.

Sensitivity analysis of coupled systems typically propagates derivatives of single-valued maps, whereas multiobjective optimization generally treats sensitivity at the system level rather than through interconnected subsystems. We develop a mathematical framework for compositional sensitivity analysis in multiobjective system design, building on monotone co-design. Subsystems are represented as design problems (DPs), interconnected through series, parallel, union, intersection, and feedback composition. We introduce tangential DPs (TDPs) to describe first-order variations around an operating point, show that they preserve the structure of DPs, and characterize their queries through graphical derivatives. To account for multiple internal realizations of a system-level operating point, tangential co-design problems compose families of local TDPs by taking a union over these realizations while retaining the original interconnection structure. We show that these compositions yield families of TDPs and establish conditions under which sensitivity analysis commutes with individual composition operations. Beyond infinitesimal sensitivity, we show that bounded tangential models of convex DPs provide tighter outer approximations of feasible finite changes, which can be propagated through co-design problems without union compositions. We conclude with several convex and nonconvex numerical examples that showcase the framework's accuracy and computational benefits.

Reinforcement learning (RL) from unit tests is the common recipe for post-training code models, but a unit test only asks whether the final program passes. We analyze what this unit-test-only reward teaches a model during debugging and find that, while it raises pass rates, it also teaches the model to edit more and to hallucinate bugs in code that is already correct. We introduce PreciseCoder, a training algorithm that augments the unit-test reward with a signal for semantically verified, localized edits, which correlates with smaller edits and higher pass rates. On 4B, 9B, and 27B models, PreciseCoder reduces the average number of edit lines from 7-15to 3 at a pass rate on par with unit-test-only training. Trained only on single-line bugs, PreciseCoder can generalize to repository-level program repair, where it resolves more issues with smaller patches. Its precision also transfers from weaker to stronger models: given the reasoning trace of a 9B PreciseCoder model, frontier models such as GPT-5.6 Sol and Qwen3.8-Max gain 4.6-12.4 points in edit precision with no additional training. Linguistic analysis of the thinking traces shows that our reward encourages the model to reason about individual lines, name the buggy ones, and form consistent bug hypotheses.

Electrotactile stimulation, as a haptic feedback method, often struggles to evoke high-quality natural sensations. Consequently, despite its practical advantages, it is generally less preferred in human-machine systems than mechanical approaches. In this paper, we propose a bio-inspired stimulation framework that generates signals modelled on afferent neural activity derived from a touch model. By producing more biologically plausible spiking patterns, the proposed approach aims to enhance the realism of perceived tactile sensations. To evaluate these stimulation patterns in a realistic task context and to quantify the influence of visual and movement cues, we conducted a series of subjective experiments under three conditions. The results indicate that bio-inspired modulation was effective: in particular, bio-inspired co-modulation of amplitude and frequency achieved a favourable pressure-irritation trade-off, increasing the probability of perceiving pressure while maintaining low irritation (pressure: 0.78; irritation: 0.24) relative to the other patterns. Visual cues and hand movement further improved feedback quality by significantly increasing the probability of reporting pressure (p$<$0.001) without a corresponding increase in irritation, and by increasing mean naturalness ratings. Overall, these findings demonstrate the potential of bio-inspired electrotactile patterns to improve haptic feedback quality, and highlight the importance of evaluating such methods under realistic, multimodal interaction scenarios.

Large language models increasingly generate Solidity smart contracts that, once deployed, can never be patched, making a single latent fault permanent and exploitable. Yet current evaluations mostly ask only whether generated code compiles. We introduce SolBehave, a benchmark for evaluating the correctness of model-generated contracts at scale, moving beyond ad hoc unit tests. Across 250 real, reproducible Ethereum contracts, we apply three complementary tests of behavioural fidelity: differential fuzzing, historical-transaction replay, and, as our core contribution, synthetic-transaction replay guided by a Dynamic Condition Response (DCR) graph that enumerates the space of legitimate function-call sequences. Differential fuzzing compares each generated function with the original on automatically generated inputs; historical replay compares the whole contract on the transactions it actually received on-chain; DCR-guided replay compares it on sequences of calls that follow its intended usage, including orderings that never occurred on-chain. We evaluate seven state-of-the-art open-weight models after one round of automated repair. Compilation success ranges from 26.4% to 92.7%, but under differential fuzzing only 31.4% of all generated function bodies that compile, pooled over models and samples, are equivalent to the original. Divergence at the contract level is the norm rather than the exception: for six of the seven models, between 13% and 31% of the generated contracts that compile diverge from the original under historical replay, and between 13% and 44% under DCR-guided replay; only Gemma-4-31B stays below 3% on both. Compilation is therefore a necessary condition for correctness and far from a sufficient one.

Gauge-equivariant graph neural networks generate ordered products of gauge links through local message passing, but the geometry of these internal structures is not captured by the receptive field alone. We investigate the statistical ensemble of closed lattice walks generated by recursive node- and edge-feature operations, independently of network training. Analytical and numerical results reveal a separation between rapid growth of contour length and much slower spatial expansion: deeper compositions generate increasingly long paths through repeated traversal of relatively compact regions. Loop expansions of fermionic observables provide a physical interpretation of this hierarchy, showing how spatial locality remains compatible with high-order dependence on nearby gauge fields. These findings identify complementary roles of network depth in extending spatial reach and enriching the representations constructed within that reach, providing a geometrical perspective on learning lattice gauge observables.

Reliable detection of pedestrians and road users is essential for the safe operation of autonomous vehicles. However, this task becomes increasingly challenging under adverse environmental conditions, such as poor weather and low lighting, where the performance of state-of-the-art pre-trained object detectors and on-board sensors tends to degrade, which limits their effectiveness and reliability. Moreover, retraining models for each operating context imposes significant constraints due to high data requirements, computational resources, and time. Considering this, we propose a novel context-aware multimodal late fusion approach that integrates a lightweight gating network within our probabilistic calibration framework. The gating network extracts environmental context and dynamically modulates the detection outputs of pre-trained object detectors based on the inferred context, without retraining of the detectors. Our methodology is evaluated on a modified KITTI dataset, in which the camera images and lidar pointclouds are augmented for varying fog levels. Experimental results show that our approach yields an approximately 14% relative improvement over non-adaptive late fusion baselines, achieving a consistent average precision of 56% across different fog levels.

Let $V_i(n)$ denote the minimum worst-case number of comparisons needed to find the $i$-th smallest of $n$ distinct elements. We construct algorithms showing $V_7(16)\le28$ and $V_8(16)\le29$, improving the previously reported upper bounds of 30 and 31. Together with known lower bounds, these results establish $V_7(16)=28$ and leave $V_8(16)\in\{28,29\}$. For 17 elements, we establish $V_5(17)=27$ and improve the upper bounds for ranks 6 through 9 to 29, 30, 31, and 31, respectively. Every new upper bound has an explicit decision-graph certificate verified by a static checker developed independently of the search. We also derive a lower bound for selection from a partially ordered state by combining the Fussenegger-Gabow weight with its order-reversed counterpart. All new upper bounds and this lower bound are formally verified in Lean 4. For $V_8(16)$, we show that a lower bound based solely on the total number of decision-tree leaves cannot establish $V_8(16)\ge29$. The algorithms were found using fixed-budget search, information-guided comparison ordering, and proof-number search. Proof-number search was decisive in finding the 28-comparison algorithm; in our experiments, refuting a comparison budget remained substantially more expensive than finding an algorithm within it.

We develop a rate-transfer method that converts finite-time Lyapunov estimates for an associated stochastic differential equation (SDE) into finite-time convergence bounds for stochastic gradient algorithms. Under generic moment and smoothness conditions, the convergence analysis of the stochastic gradient method reduces to a calculus argument verifying a Lyapunov generator inequality for the SDE. The method recovers established rates for classical stochastic gradient methods and provides a way to establish convergence rates for new optimization algorithms through calculus arguments applied to their associated SDEs.

Off-policy data in reinforcement learning with verifiable rewards (RLVR) can arise from asynchronous generation, rollout-training mismatch, and response reuse. Token-wise importance weighting avoids multiplying likelihood ratios across a response but generally biases policy gradients. We prove that, under coverage and local expressiveness conditions, full future KL credit makes the exact proximal optimizer the update's unique stationary policy, even for mixtures of older policies, without full-response weights or a fitted critic. We further establish convergence of continuous-time tabular population updates under persistent coverage and time-varying behavior policies, and a finite-sample reward guarantee for linear logistic models under additional assumptions. These results motivate Proximal KL, a practical off-policy RLVR method that separates the proximal reference from the sampling policy and replaces PPO clipping with bounded future KL credit. Across all seven matched off-policy comparisons on Qwen3-1.7B and Qwen3-4B, spanning two learning rates, response reuse, and rollout staleness, Proximal KL outperforms GRPO in micro-averaged accuracy over eight math benchmarks by up to 2.58%. A separate on-policy comparison yields comparable aggregate accuracy. In periodic-refresh experiments, Proximal KL avoids the collapses observed in PG and CISPO, exceeds GRPO's training accuracy, and reaches higher accuracy in fewer policy updates. Ablations show that token-wise correction outperforms sequence-level and unweighted controls. With token-wise correction retained, future KL feedback reduces sensitivity to the tested baseline perturbation relative to PPO clipping. A finite future horizon improves AIME24 mean@16 and pass@16 over the full tail, while fixed-policy measurements show that truncation reduces correction variance over the tested horizons.

While conventional camera-based surveillance systems and metal detectors can provide a sufficient level of security when used appropriately, they have fundamental limitations that reduce their applicability. Camera-based systems are unable to detect occluded or concealed threats and metal detectors require each person to enter through a narrow opening one at a time, leading to congestion issues and limited use in large crowds. A WiFi-based security system has the potential to be a complementary modality that is able to handle occlusions or concealment and is not hindered by strict spatial configurations. In this paper, we propose use Channel State Information estimated from WiFi-signals to detect concealed handguns. To this end we collect a dataset of aligned WiFi signals and RGB images that we call the WiFire dataset. We also propose a deep learning model called WiFireNet tailored to detect concealed handguns in WiFi signals using vision-based supervision signals. Our results show that concealed handgun detection is possible in WiFi and has promise as a complementary modality to surveillance cameras and metal detectors. We release the dataset, model and code to facilitate the further research into WiFi-based handgun detection.

Speculative Decoding (SD) has demonstrated its efficacy in accelerating large language model inference, but for Vision-Language Models (VLMs), the diverse input modalities make it difficult for a small trained drafter to achieve consistently high throughput. Self-speculative decoding reuses parts of the target model as a drafter which retains the multi-modal capability of the VLMs, yet its efficiency depends on choosing a suitable draft configuration for each prompt. We propose Single-Peak SPeculative decoding (SP$^2$ec), a training-free bandit algorithm that adapts the draft configuration during generation. SP$^2$ec exploits the single-peak throughput pattern observed in existing self-speculative methods, including SWIFT, KnapSpec, SpecVLM, and our simple cosSim. Our algorithmic design explicitly minimizes a novel notion of wall-time regret for each given query, corresponding to the empirical wall-time latency. Theoretical guarantees on the wall-time regret show that SP$^2$ec can exploit the optimal draft configuration, with negligible exploration cost. Extensive experiments on Qwen2.5-VL-7B-Instruct and LLaVA-NeXT-Mistral-7B across visual and textual tasks show consistent improvements of up to 11% over the original self-speculative methods, demonstrating the wide applicability of the proposed SP$^2$ec algorithm.

Manufacturers often need to characterize a new material, process, or product before enough operating data are available to train reliable predictive models for this new, target domain. Existing manufacturing operations contain potentially useful knowledge to accelerate this characterization, but indiscriminate reuse of data from existing source domains can cause negative transfer in learning of the new, target domain when the manufacturing domains differ sufficiently. This work proposes a reinforcement-learning method for selective knowledge transfer, where its reward scores each opportunity to transfer against supervised alternatives, so negative transfer is avoided and controlled during forward prediction instead of being diagnosed afterwards. Given limited data from a new target domain, the learned policy chooses among predictors based on individual source operations, subsets of the sources, or the supervised alternatives: fully pooled sources, or target data alone. A validation gate deploys the transferred prediction only when it improves upon the supervised alternatives. The method is designed for learning variables supporting carbon accounting, as well as performance and quality outcomes. We demonstrate the approach by quantifying the embodied greenhouse-gas emissions of parts produced via material-extrusion additive manufacturing, using data from 729 experiments spanning five materials: ABS, PETG, TPU, PLA, and recycled PLA. The response combines measured fabrication electricity with polymer-specific cradle-to-gate production factors, while 22 material, process, and toolpath features characterize each experiment. In leave-one-material-out evaluation, this method achieves a mean test RMSE of 5.2 g CO2e per part, 20.00% lower than target-only training (32.5% lower when only 5-10% of the target training data are available) and 10.30% lower than indiscriminate source-target pooling.

Vision-language models (VLMs) that segment from language must pass the location of the referent to a pixel decoder. The dominant interface compresses it into the hidden state of a single [SEG] token, a d-dimensional vector with no spatial index, and leaves the decoder, which never sees the expression, to recover the referent's extent. We show that the VLM already computes a better signal: the [SEG] token's attention over the image tokens, which is defined on the patch grid by construction. Probing a frozen segmentation VLM, we find that this attention yields 18 to 22 more gIoU than the [SEG] hidden state when read by the same probe, that it gains nothing from the image where the hidden state does, and that it is concentrated in the late layers. Building on these findings, we introduce AttnSeg, which replaces the [SEG] embedding with this attention as the decoder's prompt. A learned, signed combination of the attention heads forms a dense spatial prompt, and an attention-guidance objective supervises it directly. A two-stage recipe first trains the VLM to localize without the decoder, then trains the decoder to refine, within the budget of standard fine-tuning. Applied to Sa2VA, AttnSeg improves ReasonSeg by 6.4 gIoU and GS-Eval by 10.2 at 4B (4.2 and 9.6 over the same model fine-tuned on the same data), and by 8.8 and 11.3 at 14B. The gains reach 13 to 18 gIoU on object parts, multiple objects, and single objects, where the referent must be singled out, and are negligible on stuff regions, which the decoder already recovers unaided. Code and models are available at https://eslambakr.github.io/attseg_project_page/

Natural Language Autoencoders (NLAs) are proposed as oversight tools for language models: an Activation Verbaliser (AV) writes an English explanation of an activation vector, and an Activation Reconstructor (AR) maps the explanation back to a vector. Their value for oversight depends on a human monitor being able to tell a true account of an activation from a false one, given the Av's explanation. We formalise NLA training and NLA oversight deployment as two different referential games from Emergent Communication: training is a generative referential game, while oversight is a discriminative, multi-modal, partially-observable referential game played over a many-to-many relation between the monitored language model's activations and the accounts that are true of them. The gap between the two games predicts that AV's explanations will identify which activation they describe without discriminating its true account from closely matched false ones, since training never asks explanations to rule anything out. Using the NLA's own AR as a maximally-sympathetic surrogate monitor model, we experimentally confirm this theoretical insight on one released checkpoint: true accounts beat accounts of other activations in all 320 comparisons, but beat fluent rivals differing in a single claim only 60.2% of the time (std.err. 6.7%, chance at 50%, $n = 35$). A meaning-preserving rewrite by an uninvolved model costs only 2.1% of the distance to another activation's account, so the failure is not due to the NLA pair having developed private conventions in the AV's explanations. Our rival accounts test only transcript-level truths, which the monitored language model's conversation settles, so these results bound the easiest case and leave claims about the monitored model's own processing untested. We state conjectures that would make both results consequences of the framework.

Code-switching (CS), the alternation between languages within an utterance, is common in multilingual speech but underrepresented in pretraining corpora. Increasing CS coverage requires finding relevant utterances in large speech collections so that they can be included in training data. One way to find such utterances is to transcribe the speech and detect language alternation in the text, but this depends on obtaining accurate multilingual transcriptions. We propose a lightweight method that detects CS directly from speech. It uses frame-level language information in pretrained speech encoders to determine whether both languages are present within an utterance. Specifically, we identify language-selective hidden-state dimensions using monolingual recordings and aggregate their frame-level activations into a CS prediction. With a small CS set used to tune a decision rule, we show that the Whisper-small encoder achieves a mean F1 of $0.798$ across 14 language pairs in CS-FLEURS, compared with $0.776$ for a CNN and $0.751$ for a HuBERT linear baseline trained for monolingual language identification. Our findings suggest that pretrained speech encoders already encode the information needed to identify CS utterances, and they provide a lightweight tool for building more representative multilingual corpora.

A transformer trained on sequences whose source parameters are drawn afresh each time can learn only what the sequences share: the source family and its prior. The Bayesian predictor that knows both, which we call informed Bayes, is the unique minimizer of the training loss and is computable from the family alone; we use it as a quantitative, falsifiable reference. Where earlier work compared transformers with Bayes on sources whose contexts can be learned independently, we build Markov sources with up to 64 symbols in which observations in one context inform prediction in another: product sources with known or hidden encodings, transition rows sharing an unknown base law and depth, and a mixture of three families. Our main result is that informed Bayes predicts, from the family and prior alone, how a trained transformer behaves in four tests that go beyond the learning curve: the loss that remains under a source outside the training family (predicted 2.237 bits, measured 2.309), the doubling of the extra loss when two independent components change rather than one (0.345 predicted, 0.350 measured), the benefit of other contexts when a token appears for the first time, and the depth at which copying an earlier continuation starts to pay. On the learning curves themselves, small transformers match informed Bayes when the shared structure is visible in the tokens, and hidden structure needs more layers and training, with gaps below 0.01 bits per token for the largest models. We explain why the imitation is possible and where it must fail: informed Bayes decomposes into sufficient statistics that attention selects and accumulates and an estimator, fixed by the prior, that the feed-forward network learns; training pays the complexity of informed Bayes under the architecture's prior, and finite-memory universal prediction bounds how closely a bounded network can reach it.