Diffusion-based generative modeling has emerged as a general computational framework for simulating high-dimensional, structured data across various domains. Understanding its generative mechanisms is paramount for advancing model training and inference in a principled way. Here, we demonstrate that deterministic sampling trajectories share a universal geometric structure invariant to data modalities and diffusion model implementations. This structure can be largely explained by an exactly solvable model based on the optimal linear denoiser.
Streaming video understanding (SVU) requires models to answer user queries over continuously evolving video streams. Unlike offline video QA, SVU must handle queries whose evidence may lie in the past, appear in the current scene, or emerge only after future observations. Existing streaming video systems often rely on compact visual memories or fixed evidence-access schemes, which can lose fine-grained details and fail to adapt the answering process to the temporal intent of each query. In this paper, we propose TimeTraveler, a streaming VQA framework that performs temporal strategy planning with a Time Dictionary. Inspired by the need to search across different temporal regions of a stream, TimeTraveler stores each observed moment as a timestamp-indexed structured caption and uses this dictionary as an explicit source of queryable evidence. Given a query, TimeTraveler determines whether to recall past evidence, read the present context, or wait for future observations, and then applies a strategy-specific evidence acquisition process. Comparative experiments on streaming video QA benchmarks show that TimeTraveler improves streaming question answering by preserving fine-grained temporal evidence and selecting the appropriate temporal strategy for each query. The project page is available at https://geppa.github.io/TimeTraveler.
Retrieval-augmented generation (RAG) can surface passages that support a claim, contradict it, or fail to address it, yet existing calibrated grounding methods are typically one-sided and model only supporting evidence. We introduce Conformal Evidence Sets (CES), a two-sided framework that returns separate support and refutation sets with finite-sample inclusion coverage of at least $1-\alpha$ on each side under exchangeability; a Bonferroni union yields joint coverage of at least $1-2\alpha$. On labeled candidate pools from the contrastive VitaminC benchmark, CES closely tracks nominal coverage across $\alpha\!\in\!\{0.05,0.10,0.15,0.20,0.30\}$, with maximum deviations of $3.0$ percentage points for support and $5.4$ for refutation, and reaches essentially the same empirical coverage--size frontier as a test-tuned fixed-threshold heuristic without any test exposure. Stress tests mark the boundary of that guarantee: under FEVER distribution shift both sides under-cover, by up to $12.7$ percentage points, and support-side group-stratified recalibration substantially reduces the support gap. Scorer swaps, hard-negative retrieval, cross-dataset transfer, and a live end-to-end pilot expose a clear separation: when exchangeability holds and relevant evidence is retrieved, evidence inclusion remains calibrated even with an imperfect scorer, whereas the support/refute verdict tracks scorer quality directly. Calibrated evidence coverage and reliable verdicts are therefore distinct properties and should be reported separately.
Reproducibility is important for LoRA-adapted language models used in repeated evaluation, auditing, and decision-making. However, Split-K kernels used to introduce parallelism in low-rank LoRA multiplications can have the unintended consequence of numerical nondeterminism when partial results are combined using floating-point atomic additions. We investigate this non-determinism in vLLM across GSM8K mathematical reasoning, MBPP code generation, and Jev-style structured toxicity decisions on Civil Comments and introduce a deterministic Triton Split-K implementation. Under identical-input replay, Atomic Split-K=8 produces different FP32 shrink outputs in 97.61% of comparisons, while our Deterministic Split-K=8 shows no observed variation. Repeated atomic inference changes generated outputs in 17.30% of GSM8K and 14.96% of MBPP comparisons, sometimes altering task correctness. Structured-decision probabilities vary in 81.47% of atomic comparisons, with decision flips. Controlled tracing demonstrates how shrink-output differences propagate into the KV cache and how accumulated cache differences can change token decisions and final answers. Our deterministic Triton Split-K=8 implementation preserves parallelism, achieving 16.8-21.1% latency-based speedup over unsplit reduction with only 1.02-2.62% additional latency relative to Atomic Split-K=8.
A hedging objective reflects the risk preference of the hedger, so the conventional approach is to estimate the hedge ratio directly under the chosen objective. This paper asks whether that estimate should be used alone. We construct five objective-specific rolling hedge-ratio estimators---minimum variance, least absolute deviation (LAD), downside semivariance, lower-tail quantile, and Expected Shortfall (ES)---from the same historical stock--market returns, and combine the resulting estimates under a fixed final evaluation objective. Separate multi-output ExtraTrees models predict the within-stock-date ranks of the five candidates under MSE, LAD, downside semivariance, and 10\% quantile loss. The design uses pre-specified 60-, 90-, 120-, and 180-day windows with objective-matched rolling estimators as benchmarks. Adaptive weighting improves the corresponding loss in all sixteen objective--window comparisons. Equal weighting, which uses the same objective-specific hedge ratios without state variables or learned reweighting, already improves downside and quantile hedging. The results therefore suggest that estimates generated under other objectives can be useful even when the final hedging preference and evaluation objective are fixed, while adaptive weighting can provide a further improvement.
Automated thin-section classifiers increasingly combine plane-polarised and cross-polarised light (PPL and XPL), two complementary views central to petrographic identification. Classification accuracy alone does not reveal whether model errors reflect mineralogical relationships, how predictions depend on each optical view, or whether attribution maps identify regions that influence model decisions. Here, we examine these complementary aspects of the released LITHOS dual-stream vision transformer using outputs from an earlier evaluation of 105,802 expert-annotated grains across 25 categories. Classification errors frequently followed the label taxonomy: 72.6\% of misclassified plagioclase, microcline and sanidine grains were assigned the generic Feldspar label, while 87.1\% of quartz-family grains received a label within that family, compared with 66.6\% assigned their exact label. Replacing the XPL input with its per-channel spatial means produced a greater reduction in true-class probability than equivalent PPL replacement in 20 of 25 categories, but not in the three quartz categories. Five explainable artificial intelligence (XAI) attribution procedures generated visually distinct maps for identical inputs. In a superpixel-deletion analysis involving 30 grains from 29 thin sections, method-minus-random contrasts ranged from $-0.043$ to $0.057$, with all 41 simultaneous 95\% confidence intervals including zero. Together, these findings show that error structure, modality sensitivity and attribution behaviour capture distinct aspects of model performance that cannot be established from classification accuracy or visual inspection of attribution maps alone.
LLM agents increasingly drive systems research and rely on simulators to evaluate candidate designs at scale. Simulators are imperfect by design: their modeling limitations can distort predictions yet leave the recommended design correct, or change which design should be selected. We introduce Siris to test whether LLMs can tell these cases apart, judging from their systems understanding when a limitation invalidates the simulator's recommendation. By perturbing a simulator, Siris constructs workload pairs with the same candidates, limitation, and simulator recommendation. The perturbation changes predictions on both workloads, yet a reference simulator supports the recommendation on only one. Evaluating thirteen frontier LLMs on LLM serving design problems, we find that identifying a limitation does not ensure a correct choice, and even the strongest models sometimes abandon valid recommendations. Self-correction through a review turn helps these models restore valid choices but introduces new errors in several others. These findings motivate evaluating systems understanding through design decisions under imperfect simulators.
As Perspective API is deprecated, platforms are turning to large language models for content moderation, encoding policy as natural language that the model must interpret at inference time. Reclaimed slurs test this design directly. The same term can express in-group pride or target the community that reclaimed it, and misreading it silences the speech moderation is meant to protect. We audit four general-purpose LLMs and four moderation-specific systems, with Perspective API as a legacy baseline, on IYKYK, a corpus of reclaimed and derogatory usage of three slurs annotated by community members. Across the policy, decision, reasoning, disparity, and steering layers of this audit, we find that whether a system was purpose-built for moderation explains little about how it handles reclaimed language. No system reliably agrees with community judgment, and all fail most on language reclaimed within Black communities. What separates systems is instead how their policy is represented. Only systems whose policy is expressed in the prompt can be steered at all, and even then, explicit definitions change how models justify a decision more than how they make it. We argue that policy transparency determines whether moderation failures can be corrected, but that correcting them requires contextual judgment that definitions supplied at inference time cannot provide.
Evaluations of LLM forecasters increasingly measure how forecasts improve as richer evidence is supplied. We identify two confounds in this update-slope metric. First, outcome-aware evidence construction substantially inflates measured responsiveness: on the same models and questions, slopes are roughly four times larger than with resolution-blinded evidence ladders. Second, update slope is a change score with a prior-dependent structural ceiling, so raw comparisons can reflect differences in starting calibration rather than updating behavior. The ceiling defines only a feasible region; whether slope and prior calibration are empirically coupled depends on headroom utilization and varies across our Manifold, ForecastBench, and ACLED pools. Despite this variability, conditioning on prior calibration roughly triples the association between intermediate updating and final Brier score under both blinded and outcome-aware construction. Question identity also explains substantially more variance than model identity across the evaluated pools. We recommend building evidence ladders blind to outcomes, conditioning on prior calibration, and reporting headroom-normalized measures alongside raw slopes.
In the United States, most police departments communicate using police radio, which allows for officer-to-officer and officer-to-dispatch communication without depending on cellular network or WiFi availability. While some departments use encrypted digital radio, many continue to transmit in the clear, and are legally monitored by public citizens with `police scanners.' Because police radio is intended for internal police use and represents real-time communication, it may offer valuable insight into patterns of law enforcement activities without the potential bias and filtering introduced by a press release after the fact. In this study, we collect and examine over 14 million radio transmissions representing over 100,000 hours of speech by officers in 34 cities across the United States from July 2025 through July 2026; to our knowledge the largest scale analysis of Broadcast Police Communication (BPC) to date. We highlight both typical patterns of police communication and categorize anomalous activity that represents incidents including protests and major weather events. We present this dataset as a starting point for an array of police communication studies, conduct our own preliminary analysis, and propose several further areas of inquiry.
Open-Vocabulary Multi-Label Recognition (OV-MLR) aims to identify all queried semantic concepts present in an image. Existing methods mainly improve global image-text matching or text-side representations, leaving local prediction largely under-explored. In this paper, we revisit local prediction and reveal that it can serve as a strong discriminative branch once two properties are properly handled: spatial consistency of local visual features and compatibility-aware local competition. This perspective is motivated by local semantic sparsity: each image region is related to only a small subset of queried concepts, while compatible concepts of different granularities or types may still coexist locally. We instantiate this perspective with LoCo, a selective Local Competition framework for OV-MLR. LoCo adopts a spatially consistent visual encoder to reduce local feature entanglement, and introduces Selective Softmax to impose competition only among locally incompatible categories while preserving compatible responses. Multi-scale inference further improves local prediction at different scales. LoCo can also be ensembled with an existing global predictor to provide complementary image-level context. Experiments on four multi-label benchmarks show that standalone LoCo improves average F1-score and mAP over TagCLIP by 3.4% and 3.8%, respectively. LoCo + DualPrompt achieves the highest averages among the compared methods, with gains of 4.8% and 5.2% over TagCLIP. Our results suggest that local semantic sparsity offers a useful basis for developing more discriminative open-vocabulary multi-label recognition methods. The code is available at https://github.com/lee-plus-plus/LoCo.
Logical reasoning over natural language requires both semantic interpretation and deductive search. Large language models (LLMs) can propose plausible intermediate derivations, but they often fail to compose them into a complete proof and overpredict uncertainty. Neurosymbolic methods improve search by translating an entire problem into first-order logic and handing it to a prover, but a single bad clause can make the whole translation nonexecutable. We present Hybrid Reasoning, a three-stage pipeline that decouples proposing a proof, validating its steps, and deciding a label. An LLM proposes a structured proof tree with a raw verdict; a theorem prover validates each derivation locally, keeping only the subset it can prove; and an uncertainty-gated repair attempts to re-derive the label from the premises plus the validated derivations, when the raw verdict is UNCERTAIN. Across four benchmarks (FOLIO, ProofWriter, ProverQA, PrOntoQA) and five contemporary LLMs, Hybrid Reasoning is the top-performing method on the two hardest datasets, improving on the strongest baseline by 5.3 points on FOLIO and 4.9 on ProverQA, and reaches ceiling-level accuracy on the easiest. The repair stage operates at 78-98% precision, producing 326 correct relabellings against 19 regressions.
Due to supply constraints, replacing damaged automotive components can be expensive and time-consuming, while replacing the entire component can be wasteful when much of it remains usable. This paper presents a pipeline to investigate whether missing geometry can be reconstructed as a repair patch using paired observations of the intact models and synthetically damaged components. Noise-free color, depth, and silhouette images were produced using the camera viewpoints. Depth differences were then used to identify the potential missing regions, while multi-view voxel reconstruction was used to estimate the missing portion. A pretrained DINOv2 model was also added to validate missing candidates and determine whether pretrained visual features could improve reconstruction. We tested the pipeline on 20 automotive CAD models, each with a synthetically removed portion. Each model was tested under four different conditions using two reconstruction methods for a total of 160 trials. The predictions were compared with the missing geometry of the unchanged ground-truth using predefined thresholds of the voxel intersection-over-union over 0.85 and the absolute relative volume error less than 15%. Of the 160 experimental trials, 109 of them satisfied both thresholds. Geometry-only reconstruction passed 16 of 20 components with eight views but only 10 of 20 with four adjacent views. Further analysis showed that 24 out of the 37 failed reconstructions could be categorized as unsuccessful because extra material was incorrectly added to areas that should have been empty. The implementation of DINOv2 into the pipeline did not improve the overall reconstruction accuracy and made prediction generation slower. For 13 verified tests, printing only the repaired sections instead of the entire parts reduced filament use by 93.08% and printing time by 89.39% based on the median results.
Knowledge graphs need validated relational triples, and gold benchmarks of such triples are often built by several annotators. Most text annotation platforms focus on span tagging and compute agreement on categorical labels. They are not designed for ordinal quality grading of triples that have already been extracted. We present TripleGrader, an open-source web platform for grading subject-predicate-object triples on a 0 to 10 ordinal scale. In TripleGrader, each triple is scored independently, and peer scores stay hidden until the grader has rated it. Graders then discuss entries in threaded comments and propose revisions, which the session owner approves and publishes as new versions. For agreement, the platform computes quadratic-weighted Cohen's kappa for two raters and a quadratic-weighted Fleiss-type kappa for three or more. Graded sessions and the final benchmark are exported in common formats. We evaluated the platform in three ways. We ran 72 automated tests, including 18 workflow scenarios through the REST API. We also injected 10 error types to check whether prescribed corrections could be applied while preserving unaffected triples. On two small corpora with existing gold benchmarks, we then used the platform to curate the triples extracted by four large language models into each benchmark. Only 10% to 32% of the candidate triples could be kept unchanged, and 53% to 71% needed corrections, partly to match the benchmarks' naming conventions. Per model, 11 to 29 missing triples had to be added. On these corpora, the extracted triples therefore needed substantial human review. The source code is available at https://github.com/razz0n/triple-grader-tool.
Embedded AI is moving from perception to prediction. Robots now use world models$\unicode{x2014}$learned simulators that predict the outcome of candidate actions in a compressed latent space before any is executed. Every predicted step must be computed on board within a tight power and thermal budget, yet the energy these steps fundamentally require, and how far real hardware sits from it, have not been characterized. We derive lower bounds on the arithmetic operations and the CMOS dynamic energy of a latent state transition from information theory and Boolean circuit complexity. Profiling a DreamerV3-style Recurrent State Space Model (RSSM) shows that every rollout operation is memory-bandwidth bound and that a GPU consumes at least $5.9\times10^{10}$ times the CMOS energy floor. Transient 3D thermal simulations show that this inefficiency is also a thermal one, since sustained back-to-back planning heats a millimeter-scale silicon tile to 556 K even behind an ideal copper heat spreader. Guided by this chain from bound to bottleneck, we propose the World Model Planning Accelerator (WMPA), a chip architecture for embedded world model planning that combines near-memory computation, real-time entropy estimation, variable-precision rollout units, and copy-on-write latent scratchpads. With a design target of 40x lower power, the same tile stays near ambient at 306 K in simulation, pointing toward practical planning on thermally constrained robots.
Deployed ML models face non-stationary data and delayed labels, yet most retraining strategies rely on fixed schedules or brittle drift thresholds that ignore operational cost. In this setting, retraining is a sequential decision problem rather than a one-time threshold choice. We present DARE (Drift-Aware Retraining Engine), an offline, event-driven retraining framework that separates ingestion, monitoring, policy decisions, training, and logging through explicit schemas and deterministic replay. We compare DARE against practical heuristic retraining baselines, including fixed-interval, threshold-persistence, and action-rate-matched fixed scheduling policies. On Airline, the locked LinUCB policy achieves a higher mean F1 while substantially reducing intervention frequency relative to heuristic and action-rate-matched baselines. Additional runtime-calibrated evaluation shows that these gains persist when measured using the cumulative update wall-clock time. These results suggest that conservative sequential retraining policies can improve performance while reducing unnecessary operational overhead under delayed feedback.
Stabilizing long-horizon reasoning requires intelligent agents to sustain useful progress across successive reasoning steps and tool interactions. Such agents are typically trained with reinforcement learning (RL), where actor-critic RL uses a learnable critic network to estimate step advantages, while critic-free methods such as GRPO eliminate critic models and rely mostly on the group variants of sparse outcome-only rewards. Both reward types leave partially useful failed trajectories either implicit or indistinguishable from entirely unproductive ones, and as reasoning horizons grow, it becomes difficult to maintain consistent reasoning trajectories. In this work, we propose a \textit{plug-and-play verifier} framework for stabilizing long-horizon reasoning through explicit verification of intermediate reasoning artifacts. Specifically, the framework decouples process verification through a unified interface that extracts intermediate reasoning artifacts from agent running trajectories, evaluates them using domain-specific verifiers, and aggregates the verification scores as bounded process reward. Then, we combine our dense verified process rewards and the sparse outcome rewards through a two-stage post-training schedule. We showed that with plug-and-play verifiers, we further stabilize the reasoning steps as last-mile self-reflection and self-consistency. Typical evaluation on HotpotQA using Qwen3.5-4B achieves a superior accuracy (60.18\%), outperforming the standard GRPO baseline (57.73\%). Our results suggest that plug-and-play verifiers provide a modular approach to learn more stable and reliable reasoning behaviors over long horizons.
Vision-language models (VLMs) in medical retrieval-augmented generation must be able to say when a retrieved report contradicts the X-ray, so understanding whether they can detect such conflicts is important for safe deployment. The sycophancy mask hypothesis holds that these models internally detect conflicts but refuse to report them, concealing a detected contradiction. We introduce a three-part diagnostic that separates concealment from uncertainty: representation probing, activation steering, and two-way forced choice. Across four instruction-tuned VLMs on chest X-rays, we confirm behavioral suppression of contradiction and find that a linear probe separates matched from mismatched inputs, beyond what a simple CLIP model achieves. Steering the learned conflict direction never elicits contradiction, and forcing a binary choice does not recover one; instead, cross-modal attention to the image drops on mismatched pairs. We conclude that these models fail to contradict because the mismatch registers as hedged, gated uncertainty rather than concealed knowledge, calling for re-grounding and uncertainty surfacing. Code is available at https://github.com/mohamed-eldagla/Uncertainty-Not-Concealment/.
Knowledge of job sizes is crucial for optimizing the mean response time in single-server scheduling. However, predictions of job sizes are typically inexact and may be costly to obtain. Thus, a long-standing research question is: what is the minimal information about job sizes required to obtain near-optimal mean response time? We introduce a novel policy, \emph{Sparse Predict and Demote (SPD)}, and prove that it achieves order-optimal mean response time in heavy traffic while requiring only one-bit job-size predictions for a vanishing fraction of arrivals. SPD achieves heavy-traffic order-optimality even in situations when predictions are (i) inexact up to any fixed multiplicative error bound, and (ii) costly, i.e., the server is used to compute predictions and is thus unavailable to serve jobs while processing prediction queries. Finally, we use numerical experiments to illustrate that SPD maintains near-optimal performance beyond the heavy-traffic regime.
While multimodal large language models (MLLMs) have improved automated radiology report generation, they remain prone to hallucinations, necessitating clinically meaningful risk-control methods. Conformal prediction (CP) provides statistical guarantees, yet existing CP approaches rely on uncalibrated automated extractors and fail to account for the clinical relevance of errors. We introduce clinical admissibility criteria (CLARITY), a framework for evaluating radiology reports that explicitly distinguishes clinically consequential errors from minor deviations. To eliminate reliance on automatic judges, we collected radiologist annotations of MLLM-generated reports using CLARITY. By benchmarking CP strategies and automatic evaluation metrics against these expert annotations, we demonstrate that commonly used admissibility proxies systematically fail to detect clinically relevant errors. Our findings reveal that automated admissibility judgments are largely decoupled from true clinical admissibility.