Papers

Recursive reasoning models solve hard puzzles by applying compact, weight-tied blocks over many refinement steps. Because these blocks are reused many times, quantizing them creates a unique dynamical problem: the quantization error is incurred at every step. While 8-bit quantization (integer or float) preserves accuracy, moving to a per-tensor 4-bit format causes a systematic bias to accumulate. The ensuing drift catastrophically degrades exact-solution accuracy on Sudoku from 84.1% to 0.0% (only ~25% of cells correct). In this work, we show that this collapse is caused by activation-scaling granularity rather than bit-width or number format. Crucially, moving to per-block scaling completely restores the transition. To implement this, we apply MXInt4, a blockwise integer activation format, to recursive reasoning models. It is competitive with blockwise float formats on our tasks, while keeping integer elements and power-of-two block scales. Finally, recursion depth and reuse modulate quantization sensitivity, with the deepest architecture we test (the EqR equilibrium model) the most sensitive. Yet blockwise scaling overcomes this vulnerability, staying robust across these architectures and transferring to the open-ended ARC-AGI benchmark.

Explainability methods for time series models predominantly produce flat attribution scores: they quantify the direct influence of a feature at a timestamp by a scalar. We prove that the dominant failure mode of such methods is not the scalar format itself but a fundamental computational mismatch: existing methods compute scores via marginal conditioning or off-manifold gradients, both of which conflate direct temporal dependencies with mediated ones under autocorrelation. We also define DAG-faithfulness: an explanation is DAG-faithful if the temporal dependency graph it encodes is Markov-equivalent to the temporal directed acyclic graph (DAG) implicitly learned by the model. Particularly, we observe that standard attribution methods, specifically SHAP, are not DAG-faithful in general, and that recent time-series-aware extensions inherit the same computational limitation.

Mobile and wearable devices offer an unprecedented opportunity for continuous, passive health monitoring and active health coaching. However, the largest wearable datasets are not publicly available for research, and leading wearable foundation models trained on such datasets are rarely open-weight or come with reproducible training code. To accelerate open science in wearable health, we release OpenMyHeartCounts (OpenMHC), the largest and most comprehensive broadly accessible wearable health dataset to date, released to qualified researchers, alongside open-source implementations of recent wearable foundation models. OpenMHC, derived from over a decade of data collected through the My Heart Counts study app, includes >60 million hours of wearable data across 19 sensor channels (e.g., step count, heart rate, sleep, workouts) and up to 169 linked variables, including health, lifestyle, mood, and behavior from 11,894 consenting participants. Furthermore, we introduce a unified, open benchmark that enables standardized comparison of wearable health models across three tracks: health and behavior downstream prediction, multivariate data imputation, and time-series forecasting. We benchmark classical methods alongside recent wearable and multivariate time series foundation models. By releasing data under broad research access, alongside open-source code and model weights, at this unprecedented scale, we aim to democratize wearable health AI research and enable the community to drive open progress in this domain.

Hantavirus genomic surveillance is limited by the distribution of sequence data, non-IID source heterogeneity, and constrained expert-review capacity. We propose HantaWatch, a federated learning framework that enables laboratories and surveillance sites to collaboratively train sequence-based models without sharing raw data. HantaWatch integrates k-mer feature extraction, source-aware federated client construction, adaptive DU-FedProx optimization, surveillance-specific model selection, and prediction-only triage. Experiments on binary and multi-class tasks show that HantaWatch supports high-risk screening, outbreak-associated prediction, clade classification, and clinical-syndrome categorization while balancing predictive performance, false-negative risk, and update stability. The framework converts model output into risk scores, confidence estimates, uncertainty flags, and ranked expert-review priorities. HantaWatch therefore provides a practical federated decision-support layer for decentralized Hantavirus surveillance, supporting expert prioritization without replacing laboratory or public-health interpretation.

Integrating heterogeneous genomic and clinical modalities for joint cancer subtype classification and survival prediction remains a key challenge in precision oncology. Existing approaches suffer from three limitations: (1) they treat each modality as a monolithic feature vector, precluding fine-grained token-level interactions across modalities; (2) cross-modal fusion is typically performed through linear weighting or late averaging rather than structured token exchange; and (3) survival and classification objectives are optimized independently, missing a joint regularization signal.

A mesh of sovereign agents has no center: no shared clock, no shared model, and no coordinator to gather data or retrain. Its competence rests on each agent folding the projections its peers emit into a single internal state, online, from observations that arrive at irregular, unscheduled times, on a substrate whose weights it cannot retrain. Any one of these constraints is tractable on its own; folding optimally under all three at once is not. We ask what such a substrate must be, and prove two necessary conditions from one model of a self-evolving latent observed at irregular, exogenous times. Because the latent changes, its optimal estimator is time-varying: an adaptive timescale is necessary, and every fixed-gain filter is strictly suboptimal. And because arrivals are clock-free, the optimal estimate depends on the elapsed gap between them, which no gap-blind network recovers at any width or depth. This second condition is capacity-independent: scale cannot substitute for the missing dependence. The two conditions intersect in the continuous-time liquid class. An LSTM satisfies the first, a fixed continuous-time filter the second, and a multi-timescale liquid network both. Synthetic experiments confirm each: the network attains the timescale, and the separation is computed exactly. The characterization is necessary, not sufficient, and binds fixed-weight substrates: a network free to retrain reaches the class by other means. Proved per agent, the necessity binds every agent of a mesh, a structural condition on mesh intelligence.

Operations Research (OR) provides a rigorous framework for high-stakes decision-making, but effective OR modeling requires substantial domain knowledge, mathematical abstraction, and solver expertise. Recent LLM-based systems automate parts of this pipeline, yet remain limited by low accuracy on complex problems, opaque outputs, and narrow solver support. We propose COOPA (COoperative OPerations Agent), a modular LLM-agent architecture for interpretable and scalable OR decision support. It combines three components: iterative confidence-based modeling, which generates multiple candidate formulations, self-evaluates them across modeling dimensions, and selects one using a max-min confidence criterion; element-level provenance and confidence explanations, which link variables, parameters, constraints, and objectives to quoted source text and provide an audit trail for human verification; and multi-solver routing to specialized optimizer agents for different OR problem classes. Across three OR benchmarks, eight LLM backbones, and four baselines under identical conditions, COOPA achieves the best macro-average accuracy on six of eight backbones and improves over the strongest baseline by up to 6.7 percentage points. A within-system ablation isolates the contribution of iterative confidence-based modeling, while additional analyses and case studies illustrate the value of source traceability and multi-solver dispatch.

Physical or operational constraints often impose communications limitations on autonomous agents. Such limitations complicate monitoring or multiagent coordination. Even when strong communications are absent, some information may still be available. The remainder of the relevant agent state may be reconstructed via estimation. The actions taken by an agent are a potential source of information – as the agent interacts with the environment, these actions may be observed even in the absence of explicit communication. We investigate using actions to estimate the state of an agent, using reinforcement learning to develop policies which make the estimation problem more tractable. Policy observability is encouraged through the training reward and is analyzed using simulation of the trained agent. In an aircraft tracking problem a policy with enhanced observability is found that has minimal impact on nominal task performance.

We present Qwen-Image-2.0-RL, a post-training pipeline that applies reinforcement learning from human feedback (RLHF) and on-policy distillation (OPD) to improve both the visual quality and instruction-following capability of the Qwen-Image-2.0 diffusion model. To provide reliable reward signals, we construct task-specific composite reward models by fine-tuning vision-language models with a pointwise scoring paradigm and chain-of-thought reasoning. For text-to-image generation, the reward models cover alignment, aesthetics, and portrait fidelity dimensions. For image editing tasks, the reward system addresses instruction-following accuracy and face identity preservation. Building on this reward system, we develop a scalable GRPO-based RL training framework, incorporating a hybrid classifier-free guidance (CFG) strategy to preserve pre-trained knowledge, prompt curation via intra-group reward range filtering, and per-category reward weight calibration. To merge the task-specialized RL policies for T2I and editing, we propose on-policy distillation as the final training stage, which consolidates multiple teachers into a single student model through trajectory-level velocity matching. Extensive evaluation shows that Qwen-Image-2.0-RL achieves 57.84 overall score on Qwen-Image-Bench (+2.61 over the base model), Elo ratings of 1193 in text-to-image arena (+78) and 1349 in image edit arena (+93), demonstrating consistent gains in aesthetic quality, prompt adherence, and editing accuracy.

While many explainable AI (XAI) methods have been proposed, most are not designed for time-series forecasting models and often rely on the implicit assumption that timestamp features are independent. This assumption ignores the fundamental property of temporal dependence and can lead to explanations that violate the sequential and causal structure of the data. We introduce KARMA, a method for explaining time-series predictors by constructing a Markov surrogate model that captures the temporal dependencies learned by the predictor. Our approach revolves around three main aspects: identifying the minimal history length $K$ that is predictively sufficient for the model, estimating the best-fitting $K$-order Markov transition kernel from the discretized history space, and a five-level global explanation hierarchy that can be derived from the Markov transition kernel, which we illustrate using real-world weather data (Beijing PM 2.5). We also certify using complex synthetic data with known true causal edges that KARMA (i) recovers the data causal structure as learned by the model via a controlled experiment and (ii) identifies temporal dependencies better than established attribution methods such as TimeSHAP.

We introduce a categorical framework called ODYSSEY for constructing verifiable, local truth-preserving foundation models as compositions of foundries: building-block architectural components that specify a cover of local contexts, local representation families, restriction maps, gluing rules, obstruction policies, update obligations, and human-facing views. A foundry is an organized sheaf of knowledge that carries within it an argumentation component. Concrete foundries are built from generic foundries such as evidence/argument, operational decision, institutional/financial, market meaning, scientific challenge, research-program, assistant-build, and evaluation-harness foundries. Universal Foundry Learning (UFL) formalizes foundry construction as a composition of left and right Kan extensions, with left Kan extension rolling local artifacts into candidate foundries and right Kan extension enforcing the restriction, gluing, obstruction, and argumentation conditions required for promotion. Foundry SQL (FSQL) is a small typed query surface for slicing maintained foundry artifacts that uses TICKET (Topos Integration using Causal Kan Extension Transformers) certification for admitting external or pre-built models into durable ODYSSEY state. ODYSSEY is fully implemented and tested across a wide spectrum of concrete foundries, showing that the same categorical machinery supports domain construction, artifact replay, sheaf diagnostics, grounded Toulmin/local-LLM scrutiny, residual-obstruction ledgers, and optimized TICKET-compatible causal-claim extraction across heterogeneous sources. This paper is to be presented as a 2.5 hour tutorial at ICML 2026. The tutorial home page is at https://bit.ly/4ajS0nA.

Reinforcement learning from human feedback (RLHF) in production does not always have a synchronous reward signal. Code-execution verifiers, slow judge ensembles, and queued human review can return several gradient steps after the rollout that produced them, breaking the synchronous-reward assumption underlying standard PPO. We address this gap with Retroactive Advantage Correction (RAC): each pending slow completion is queued, aged through a non-negative kernel, and reinjected as a clipped residual into the next optimiser step's advantage. We prove that under an unbiased clipped importance ratio, the cumulative RAC correction is exactly unbiased when the effective delay kernel reinjects all of its mass, and carries a bias linear in the unreinjected fraction otherwise; at the no-delay identity kernel it reduces to V-trace. On a tabular Markov decision process (MDP) proof-of-concept, RAC reduces the closed-form policy bias by up to 47.9x at the two-slow-channel configuration, beating wait-for-slow at lower wall-clock cost. RAC integrates with PPO and GRPO through a two-line reward-manager patch.

Accurate assessment of tumor proportion score (TPS) in non-small cell lung cancer (NSCLC) is critical for treatment planning and prognosis. Key challenges include the tedious manual work required to annotate each slide, combined with the limited number of experts certified for this task. Multiple instance learning (MIL) has proven to be an effective approach for predicting TPS scores at the slide level; however, existing methods struggle with non-expressive (zero class) images. Our approach involves two models: (1) an embedding-extraction and multiclass-classification network that captures the histopathological features of individual patches, and (2) a MIL model that aggregates these embeddings to predict zero-inflated beta (ZIBeta) parameters representing the overall TPS probability distribution for the entire slide. Using only slide-level TPS scores as labels, we demonstrate how this end-to-end framework can leverage a novel distribution-based architecture to improve prediction accuracy and explainability. ZIBeta modeling significantly outperforms baseline linear and ridge regression while capturing expected accuracy through distribution concentration.

Reward models for Reinforcement Learning from Human Feedback (RLHF) pool preferences across thousands of annotators and fit one global affine calibrator, collapsing raters with systematically different rating-scale offsets and slopes into a single average-rater fit that does not match any individual annotator. PEBS is a per-rater empirical-Bayes shrinkage estimator: it fits per-rater affine calibrators on a held-out slice of each annotator's ratings and applies Morris-James-Stein empirical-Bayes shrinkage toward the population mean, in closed form and without retraining the reward model. On PRISM, PEBS reduces within-user held-out RMSE by 8.58% over the pooled population-slope baseline. The procedure replicates on PluriHarms harm ratings (Qwen-2.5 base, in-family) with a +9.66% RMSE reduction over the same population-slope baseline. PEBS is a closed-form post-hoc estimator for annotator-specific affine calibration in RLHF reward modeling; it leaves the reward base model unchanged and estimates only the rater-level map used at inference time for new ratings.

This paper formulates frame-level freeway risk assessment as a multi-agent scene graph-level binary classification problem, where each video or trajectory frame is labeled risky if any TTC- or PET-based conflict violates a specified severity threshold. We construct a relation-aware graph per frame with vehicles as nodes and two interaction types as edges: same-lane (longitudinal) and adjacent-lane (lateral), augmented with physics-informed edge features aligned to rear-end and lane-change conflict mechanisms. Building on a structured benchmarking suite of non-graph models and graph baselines, we propose HIA-GAT, a dual-stream heterogeneous graph attention network that processes longitudinal and lateral interactions through dedicated attention pathways and fuses them via a conflict-type-aware gating mechanism with event-level gate supervision derived from SSM conflict attribution. Experiments on the NGSIM I-80 and US-101 freeway datasets across nine TTC and PET threshold configurations show that HIA-GAT achieves the best average risk-ranking performance (AUC 0.835 on I-80 and 0.867 on US-101), with the largest gains on PET-only (lane-change) settings where relational structure is essential. Beyond accuracy, the learned gate provides interpretable per-vehicle attribution of dominant conflict type, supporting actionable, real-time freeway safety monitoring. We show that graph structure is critical for modeling lateral conflict risk, while longitudinal risk can often be captured by non-relational aggregation.

Prompt injection is the top security risk for LLM-integrated applications, yet every defense proposed so far has been broken. We prove this is not a coincidence: in shared-embedding architectures that lack enforced control-data separation, perfect prompt-injection prevention is mathematically impossible. We formalize prompted systems as Prompted Action Models whose outputs include control-authoritative actions: refusal decisions, tool authorization, policy routing, and memory writes. We define Semantic-Faithful Control (SFC), the property that such behavior depends only on the meaning of untrusted input, not on how it is encoded. We then prove SFC is unachievable within the shared pipeline, via three results: a provenance-recovery impossibility (shared representations make trusted and untrusted content statistically inseparable, bounded by total variation distance); control-path exposure (untrusted tokens enter control-relevant computation through the same attention value-aggregation that determines outputs); and a finite-coverage invariance gap (finite training cannot certify invariance over infinite semantic-equivalence classes). We ground each quantity in measurements on production tokenizers and models. The result is structural, not a gap in current defenses. It mirrors the code-data confusion in Von Neumann machines that gives rise to buffer overflows, a vulnerability class that took decades of layered defenses (DEP, Write-XOR-Execute, ASLR, stack canaries, and ultimately memory-safe languages) to contain, because no single mechanism sufficed. The implication is the same: prompt injection cannot be eliminated by better in-pipeline classification or alignment alone. It requires architectural separation of instruction and data channels. We identify the root cause and the class of solution it demands.

Generative models have achieved remarkable success in data synthesis, though recent advances driven by increasing model scale have introduced challenges in computational cost and efficiency. Quantum machine learning offers a promising alternative, representing complex data distributions using compact, highly expressive models. Here, we propose QDiffusion-TS, the first quantum generative diffusion model for time series synthesis, and validate it on the IQM quantum processor. The framework extends a classical diffusion architecture by replacing feed-forward components within the denoising transformer with quantum neural networks, yielding a hybrid quantum transformer that reduces the number of trainable parameters in each replaced component by nearly three orders of magnitude. Evaluated on financial time series from Apple and Amazon, the model generates synthetic data that more accurately reproduces the real distributions, reducing Wasserstein distance by approximately 44% relative to its classical counterpart across both datasets. In a downstream forecasting task, augmentation with the generated data improves predictive performance by up to 71% in RMSE over a baseline trained solely on real data. These results show that quantum enhanced architectures can consistently match and frequently surpass classical performance with substantially fewer parameters, establishing a practical framework towards more efficient and scalable data-driven generative modelling.

Fairness measurements in the form of disaggregated evaluations often rely on demographic signals that are legally constrained or culturally sensitive. Race and ethnicity signals are among the more difficult signals to curate and use for this task. This paper presents Privacy-Preserving Probabilistic Race/Ethnicity Estimation (PPRE) as a method for enabling fairness measurements with respect to race/ethnicity for U.S.\ LinkedIn members in a privacy-preserving manner. PPRE applies privacy technologies (specifically: secure two-party computation, differential privacy, and additive homomorphic encryption) on top of two race/ethnicity demographic signal sources (the Bayesian Improved Surname Geocoding estimator and a sparse golden survey set of self-reported demographics) to power a fairness measurement solution with respect to US-based race/ethnicity demographics. We detail its privacy guarantees and demonstrate its application on candidate- and viewer-side fairness measurements. We close with a transferable framework for institutions seeking to implement similar privacy-preserving measurement infrastructure.

Multi-token prediction has been shown to increase data density during training, improve downstream text-generation quality, and serves as the defacto approach for self-speculative decoding. Existing foundation and open source models that use MTP heads commit to a static tree-based attention topology throughout the entire generation sequence, meaning the speculation depth, and thus the compute required during verification, stays constant regardless of the context. This is fundamentally misaligned with the entropy patterns of natural language where low-entropy regions often support reliable multi-step drafting, while high-entropy regions require more conservative speculation. To address this, we propose Entropy-guided Multi-Token Prediction (EntMTP), a training-free scheduler that toggles between tree-based attention topologies from a set of task-specific pareto-optimal trees conditioned on a running estimate of local generation entropy. By matching speculation depth to context predictability, EntMTP maximizes expected accepted-token throughput across the full distribution of generated text without sacrificing generation quality. When evaluated across Humaneval, ShareGPT, GSM8k, and Litbench benchmarks, EntMTP consistently achieves a 1.15x speedup against Hydra and peak speedup of 1.36x against Medusa baselines respectively.

The variations in vocal effort range (e.g. whisper, soft, neutral, loud, shout) alter production and speech acoustics, reducing intelligibility and limiting the robustness of any subsequent speech technology. Classification is challenging since effort lies on a continuum, adjacent categories are easily confused, and labeled data remain scarce. Prior SSL approaches with wav2vec2, HuBERT, and AST improve performance on the AVID corpus but still suffer from boundary errors. In this study, we introduce WavLM for the first time in vocal effort classification and benchmark it against wav2vec2 and HuBERT. To address data scarcity, we conduct a systematic study of augmentation strategies, covering RIR convolution, additive noise, time masking, speed perturbation, band-limiting, MixUp, and CutMix. Augmentation consistently improves WavLM, with gains ranging from +0.6% to +1.8% absolute. We further propose Gaussian-neighbor soft labels, which further reduce near-boundary confusions by modeling the vocal effort continuum. Our best system, WavLM-BASE with gradual unfreezing, augmentation, and Gaussian-neighbor soft labels, achieves 78.2% mean accuracy, establishing a new state-of-the-art on AVID.

Social media popularity prediction aims to forecast the future reach or influence of online content from early-stage observations. Accurate prediction enables key downstream applications, such as advertising optimization and strategic content planning by users, creators, and platforms. Despite substantial progress, existing popularity prediction works often fail to jointly consider multimodal content and temporal social interaction signals. Moreover, the literature remains highly fragmented across datasets, modalities, observation windows, prediction targets, and evaluation protocols. This fragmentation prevents fair comparison and obscures a systematic understanding of how textual, visual, temporal, and interaction-based signals jointly shape popularity dynamics. To address these challenges, we introduce MMG-Pop, a Multi-modal Graph-based Popularity Prediction benchmark, which unifies datasets, modalities, temporal interaction signals, and representative baselines under a standardized evaluation protocol. Furthermore, we propose MMG-PopNet, a unified multi-modal graph-based network that jointly models the aforementioned multi-modal signals and graph-structured social interactions. Extensive experiments on MMG-Pop, comprising four datasets across Bluesky and Reddit platforms, demonstrate the superior performance of MMG-PopNet and yield new insights into cross-platform training generalization, multi-task prediction benefits, multi-modality contributions, and LLM prediction limitation. These findings establish a unified foundation for future research on social dynamics modeling and intervention under heterogeneous modalities and socially-aware agentic ecosystem paradigms.

Speech emotion recognition (SER) often relies on hard consensus labels that collapse annotator disagreement. We study distribution-based supervision for 9-class SER on MSP-Podcast 2.0 using a WavLM-Base multitask model for categorical emotion and dimensional VAD. Hard-label training is compared with targets from primary and merged primary–secondary annotator vote distributions. Distributional objectives improve alignment with human vote distributions, reducing JSD/KLD relative to hard-label training. Analysis shows that hard supervision partly benefits from assigning ambiguous utterances to the residual Other class, whereas distributional supervision redistributes uncertainty across emotion categories. Entropy-stratified evaluation shows that high-ambiguity utterances remain challenging, but distribution-based supervision better captures perceptual uncertainty. These findings support moving beyond hard labels toward targets that reflect listener disagreement.

Large Language Models (LLMs) possess broad conceptual knowledge acquired through large-scale text pretraining, yet their potential to supervise models in other modalities remains underexplored. In this work, we propose LaViD–Language-to-Visual Knowledge Distillation–a simple and effective framework for transferring high-level semantic knowledge from a language-only teacher to a vision-only student model. Instead of relying on paired multimodal data, LaViD elicits conceptual signals from an LLM by prompting it to generate multiple-choice questions (MCQs) that probe semantic distinctions between visual classes. Each class is mapped to a soft label distribution over these MCQs, forming a rich conceptual signature that guides the student through an auxiliary distillation loss. Notably, despite using a language-only teacher without access to image data, LaViD consistently outperforms recent methods like MaKD that distill from vision-language models across multiple fine-grained benchmarks. It also achieves competitive or superior performance compared to state-of-the-art visual distillation methods such as DKD and MLKD, with further gains when combined with logit standardization. On the Waterbirds dataset, LaViD substantially improves worst-group accuracy, demonstrating enhanced robustness to spurious correlations with distillation. Code is available at https://github.com/lliangthomas/lavid.

Accurate prediction of bottom-hole pressure (BHP) and CO2 plume migration is essential for safe geological carbon storage, yet practical simulations often rely on truncated domains where artificial boundaries distort pressure diffusion and CO2 saturation footprints. In this study, we evaluate how boundary-condition fidelity affects BHP and CO2 plume prediction by comparing ten reduced-domain boundary treatments against full-domain reference simulations in homogeneous and heterogeneous reservoirs. We test uniform pore-volume multipliers, transmissibility modifiers, corner-adjusted pore-volume corrections, layered corrections, and gradual modifiers using BHP RMSE, NRMSE, peak pressure deviation, and plume Intersection over Union (IoU) as performance metrics. Our results show that conserving corner pore volume is the most important requirement for truncated-domain modeling. We find that uniform treatments which neglect corner storage generate large pressure errors, with BHP RMSE of 362 to 382 psi in the homogeneous model and 250 to 304 psi in the heterogeneous model, and yield plume IoU values near 0.80 to 0.84, indicating roughly 16 to 20% of the combined plume area is misrepresented. Corner-adjusted scenarios substantially reduce pressure errors and raise plume IoU above 0.94, but we observe that transmissibility correction is not universally beneficial. In homogeneous reservoirs, uniform transmissibility adjustment improves pressure fidelity; in heterogeneous reservoirs, it can over-restrict flow across variable-permeability boundary faces, increasing BHP error and contracting the predicted plume. We find the gradual modifier with transmissibility correction provides the most consistent performance, achieving BHP NRMSE below 3.7% and plume IoU above 0.97 in both reservoir types.

Activation patching is the primary tool in mechanistic interpretability. It attributes causal responsibility for a model behavior to each of its individual components by estimating its natural indirect effect (NIE). Re-deriving the activation patching estimand from causal mediation analysis, we find that the NIE does not solely capture the causal effect through the specific component. It also contains interaction effects (INT) that measure how much the component's causal effect itself depends on the state of other components in the model. A natural response may be to try to eliminate INT by adjusting the estimator or unit of analysis, but each of these potential remedies has predictable failure modes. We demonstrate these failure modes in the GPT-2 IOI circuit; components whose causal importance is conditional on the state of other components are either invisible or artificially inflated, and INT variance explains the previously documented instability of faithfulness scores. We prove that INT scales with the distance between clean and patched component activations, is negligible when the model is locally affine, and decomposes combinatorially into pairwise and higher-order group interactions. Despite its inevitability, INT is not a nuisance to be eliminated, but rather a diagnostic for interpretability studies. Its individual and group-level magnitude and sign signal when causal conclusions are prompt-dependent, and when greedy NIE-based component ranking will miss mechanisms only discoverable through combinatorial search.

We present a first study of a diffusion-based approach to accelerated sampling of the $N_f = 2$ lattice Schwinger model. Our work is inspired by recent and growing successes in developing such generative models for ensemble generation in LFT to overcome the well-known critical slowing down problem. We train a U(1)-equivariant score-based generative model to sample gauge link configurations from the marginal Schwinger model. By computing model likelihoods, we obtain unbiased estimates for observables that closely match those produced by MCMC simulations. We also demonstrate improvement over HMC as measured qualitatively by a reduction in topological freezing near critical parameters.

Robots trained on real world data tend to be imprecise, slow, and brittle to perturbations. Improving these policies with reinforcement learning (RL) is an appealing alternative, but this process often requires expensive training in the real world. Performing policy improvement in simulation instead provides a far cheaper alternative, but unconstrained RL in simulation can exploit contact and dynamics mismatches, resulting in unsafe behaviors that do not transfer to hardware. Common forms of regularization can furthermore limit improvement by overconstraining to an imperfect behavior prior. In this work, we propose Support-Constrained Off-Domain REinforcement (SCORE), a real-to-sim-to-real framework that constrains RL in simulation to the support of a generative policy pretrained on real data. We instantiate this constraint through flow steering, restricting SCORE to actions the base policy can already produce, which ensures transferable behaviors while maximizing policy improvement. Improving a policy with SCORE requires minimal effort: it learns from sparse rewards, avoids distillation, and leaves the base policy untouched. Across eight real-world dexterous multi-fingered robotic manipulation tasks, SCORE improves average success rate from 37.8% to 89.9%, compared to 59.5% for the best baseline, and reaches success in 36.8% fewer steps than the base policy. Ultimately, through extensive experiments and ablations, we show that simulation can substantially improve real-world manipulation policies when policy optimization is appropriately constrained, introducing a new paradigm for real-to-sim-to-real policy improvement. Videos and code are available at https://weirdlabuw.github.io/score/.

Large language model (LLM) agents operate over long, multi-session interactions in which facts change: a user moves, a price updates, a plan is revised. Acting correctly requires using the current value of a fact and discarding values that have been superseded. We isolate this ability on real conversational data and show that it is a distinct, unsolved failure. On the knowledge-update subset of LongMemEval, replacing an agent's full context with a bounded, self-maintained memory drops accuracy from 92% to 77% even on a frontier model (gpt-5.4), a gap that is statistically significant (paired McNemar p<0.005) and persists across model scale while full-context accuracy saturates near 92%. The bottleneck is therefore memory maintenance, not comprehension, and is not closed by a stronger model. We then ask whether this is merely an undersized memory, and find it is not: as the conversation grows 24x, accuracy falls further (from 68% to 28%), and granting the agent proportionally more memory yields no detectable recovery (28% to 28%, n=25). The failure scales with the length of the conversation, not the compression ratio. We release Supersede, an open reinforcement-learning environment (on the verifiers / prime-rl stack) that turns this measurement into a training signal: agents are rewarded for answering from the current value and penalized for stale ones. Finally, we close the loop and show the gap is trainable: GRPO fine-tuning a small open model (Qwen2.5-3B) on this environment nearly doubles its held-out supersession accuracy on real, unseen conversations (9.0% to 16.7%, a single run), along a monotonic checkpoint curve indicating the learned policy, not the harness, carries the gain. To our knowledge this is the first trainable environment whose reward targets temporal fact-currency, and the first evidence the supersession gap can be trained down, not only measured.

The global minimum-variance portfolio (GMVP) is the canonical decision built from an estimated covariance matrix, yet covariance estimators are universally evaluated by matrix-norm loss, which is not the object the decision depends on. We characterise exactly how covariance-estimation error maps into GMVP suboptimality. We prove an exact regret identity and a non-asymptotic bound showing decision regret depends on the estimation error only through its action on the portfolio weights, scaled by portfolio concentration and the conditioning of the true covariance. From this we derive the decision geometry: GMVP regret is invariant to a (p-1)-dimensional projection of the p^2-dimensional error matrix, with invariance to the covariance-scale direction as an exact special case. We then apply the framework to heavy-tailed returns (tail index kappa in (2,4)), establishing the regret convergence rate implied by the centred operator-norm rate, and confirm the theory on a skew-t/t-copula simulation design with pre-registered analysis. The decision-focused advantage is a sharper constant and a concentration discount rather than a faster rate; we report an honest high-conditioning boundary of the rate prediction. The results complement recent decision-focused learning approaches by supplying the exact estimation geometry and consistency theory they lack.

We consider the cubic nonlinear Schrödinger (NLS) equation on two-dimensional flat tori with varying aspect ratios. In this formulation, the choice of aspect ratio governs the Fourier resonance structure, so rational and irrational geometries can exhibit different high-frequency cascade behaviors. We present a geometry-conditioned Fourier neural operator (FNO) for the cubic defocusing NLS equation, where the input consists of the real and imaginary parts of the solution together with the aspect-ratio parameter \(\omega^2\). The model is trained to approximate the one-step solution operator and is evaluated on unseen trajectories generated from random-phase initial data using Fourier pseudospectral method. Our numerical experiments show that the learned operator captures the main solution dynamics on both tori and reproduces the distinct Sobolev norm behavior of the two geometries, with stronger \(H^2\)-growth on the rational torus and more constrained behavior on the irrational torus, consistent with the findings of \cite{hrabski2021energy}. We perform ablation studies to examine the roles of retained Fourier modes, activation functions, Fourier-layer depth, and explicit geometry conditioning. The results indicate that including $\omega^2$ improves long-time predictive accuracy, especially for the rational geometry, and supports the use of geometry-aware neural operators for learning spectral-transfer phenomena in nonlinear dispersive partial differential equations.