Hazard and Operability (HAZOP) studies safeguard industrial processes against catastrophic accidents. Large language models (LLMs) increasingly assist HAZOP drafting, yet evaluations rely on lexical overlap. Because LLMs internalize human heuristics, prompts may trigger biased judgments. Whether frontier LLMs withstand cognitive perturbations in safety-critical workflows remains unknown. Here we show across 420 trials that three frontier LLMs resisted 96.8% of cognitive manipulations, rendering overall susceptibility statistically inconclusive. Contrary to assumed broad fragility, models rejected confirmation and availability cues. However, safety-critical risk downgrades emerged under low anchors and gain frames. Thus, automated hazard analysis requires disciplined human-in-the-loop validation.
Vision-Language Models (VLMs) are increasingly adopted as end-to-end driving planners, yet they suffer from cross-city domain shift. Existing methods treat this shift as a single quantity addressed by uniform alignment or single-site editing. We challenge this view on two counts: (i) domain shift in VLM trajectory planners is inherently structured across modality and layer, and (ii) a locate-then-repair recipe matched to this structure outperforms every uniform-objective baseline. We analyse domain shift via activation patching and uncover two regularities. (F1) Modality axis. Image and text tokens exhibit distinct shift patterns in each VLM. (F2) Layer axis. The layer-wise shift partition is architecture-dependent and dataset-stable. These findings recast domain shift as a two-dimensional tensor indexed by modality (image vs. text tokens) and layer, motivating a simple principle: locate where shift occurs, and repair only there. We instantiate this principle in MoLaRx (Modality-Layer Repair), a locate-then-repair framework. The Causal Importance Score (CIS) estimates the domain shift as a modality-layer tensor and partitions the layers into functional zones. CIS-Guided Selective Repair (CISR) then applies a matched operator (Maximum Mean Discrepancy (MMD) / low-rank adaptation (LoRA) / freezing) to repair each zone, with the composition adapting automatically to each architecture. We evaluate MoLaRx on three architecturally distinct VLMs (LFM2-VL-1.6B, Qwen2-VL-2B, and PaliGemma2-3B) across two cross-city benchmarks (nuScenes and Argoverse 2). MoLaRx attains the lowest cross-city trajectory-error gap among adaptation methods in all six main-table settings, while using 50-60% fewer trainable parameters and 8.4-10.9 ms lower inference latency. In contrast, uniform MMD-only adaptation collapses in all six settings, highlighting the importance of structure-guided repair.
Pre-training is where every foundation model begins; bringing recursive self-improvement (RSI) there would let AI improve the process that creates AI. Yet public RSI work has so far targeted only post-training, which starts from an existing base model, runs trials costing GPU-hours, and returns verifiable scores. Pre-training breaks all three conditions at orders-of-magnitude higher cost: a run consumes hundreds of thousands of GPU-hours over months and cannot be repeated; its objective, post-training headroom, cannot be read off any single score; and what must improve is not a model but the whole production system, from data pipelines and training infrastructure to the teams that run them. To our knowledge, we are the first to apply RSI to an entire frontier pre-training run from scratch and to report it in detail. RSI does not stop at frontier pre-training, but it changes form: since no end-to-end loop can close over a single unrepeatable run, we decompose it into smaller loops that can, each repairing a step the full run breaks. The result is 801-30B-A3B-Base, a 30B-A3B MoE trained from scratch on 11.8T tokens in three months by a newly formed team. We release it as a base model; the lightweight SFT reported here (6B tokens, no RL) only probes post-training headroom. As a base model, 801-30B-A3B outperforms Qwen3-30B-A3B and Nemotron-3.5-Lightning-30B-A3B, trailing only the larger Qwen3.5-35B-A3B, and keeps this position under identical SFT; with lightweight SFT alone, it achieves the highest average on four agentic benchmarks against fully post-trained models, including Qwen3.6-35B-A3B. Since post-training RSI presupposes a strong base model, bringing RSI into pre-training acts on the foundation every later stage builds on. We open-source 801-30B-A3B-Base and the lightweight SFT data at https://github.com/GAIR-NLP/801-30B .
Anomaly detection plays a critical role in securing the Internet of things (IoT). Although existing federated principal component analysis (PCA) methods can learn the subspace of normal samples without sharing raw data, they often suffer from feature redundancy that may degrade detection performance. To address this limitation, we propose a novel and effective federated sparse Grassmann Newton (FedSGN) learning method for IoT anomaly detection. Technically, our FedSGN leverages the $\ell_{2,0}$-norm constraint on the global projection matrix to eliminate redundancy and introduces an orthogonality penalty to preserve the intrinsic structure. Moreover, we develop an augmented Lagrangian solver integrated with a semismooth Newton scheme to enable parallel client updates and Grassmann projection for robustness. Numerical experiments demonstrate that the proposed method achieves favorable overall detection performance while reducing communication overhead. The source code is publicly available at https://github.com/xianchaoxiu/FedSGN
With a solid theoretical foundation and pretraining on large-scale datasets, diffusion-based methods have achieved controllable generation leveraging text and image conditions, and generating excellent stylized images is possible using diffusion models. However, previous methods face significant challenges. First, because diffusion generation is iterative, the common approach of learning implicit style representations via cycle-consistency strategies is too costly and not well-suited to diffusion, even for LoRA-based methods. Second, decoupling style is difficult in the latent space of diffusion models. Representations of attributes are coupled in the latent space of diffusion models; thus, extracting style attributes with diffusion models usually suffers from the issue of content leakage. Furthermore, because style is hard to represent in natural language, capturing style using text prompts or textual inversion is deficient. In this paper, we propose a method for achieving disentangled stylized image generation through IP-cross-attention rearrangement, based on our experimental observations. We also propose negative content guidance, inspired by classifier-free guidance, to reduce content leakage in stylized image generation.
Image-text alignment models such as CLIP are typically trained with contrastive learning, where unpaired examples are treated as strict negatives. While this assumption introduces noise by penalising semantically related pairs, existing solutions often rely on heuristic similarity thresholds that lack statistical grounding and sensitivity to dataset-specific noise distributions. In this paper, we propose Permute-then-Adapt (PTA), a framework that leverages efficient weak models to identify missed positives while controlling Type-I error relative to a conditional permutation reference, up to calibration approximation error. Unlike standard distillation or arbitrary thresholding, PTA estimates a random-pairing reference distribution from the weak teacher's similarity scores and uses its upper tail, determined by $\alpha$, to select auxiliary positives. The calibrated positive set is then trained against by a single multi-positive contrastive objective: each anchor maximises the total probability mass it assigns to its positive set. Offline calibration with the weak model adds negligible training overhead. Experiments show gains over heuristic soft-label approaches in object recognition, cross-modal retrieval, and data efficiency. Project repository: https://github.com/JinhaoLee/Permute-then-Adapt.
Stroke-based rendering is often optimized as a pixel-level reconstruction task. This over-reliance on global error reduction not only degrades strokes into fragmented curves due to a lack of structured planning, but also allows locally defective primitives to be concealed by overlapping layers. To address these issues, we introduce a model-guided framework for stroke-audited rendering. Our framework uses Quadratic Bezier Splatting to analytically model continuous curved strokes with efficient differentiable alpha compositing. Built on this renderer, a feedforward multi-scale stroke decoder provides a structural initialization with learned stroke-scale allocation, while test-time stroke optimization supplies the reconstruction capacity needed for high-fidelity refinement. We further propose a single-stroke audit mechanism that explicitly evaluates local stroke quality, preventing defective primitives from surviving optimization through layer-wise compensation. Compared with existing methods, our approach achieves coherent layouts, improves PSNR by up to 5.46 dB, yields a 16% more compact representation, and accelerates combined renderer forward and backward computation by a factor of 220.3.
Abductive reasoning is the inference of plausible explanations for observed evidence. Evaluating this capability requires more than a single task: a system may need to generate an explanation, discriminate among competing hypotheses, revise its hypotheses as evidence arrives, or decide what evidence to seek next. Yet these capabilities are typically evaluated through isolated benchmarks with different task formulations and scoring procedures, so claims that a model is ``good at abduction'' may reflect only a narrow aspect of this capability. We introduce AbductionBench, a unified framework whose taxonomy characterizes abductive tasks. We instantiate this framework on 27 datasets spanning ten domains and evaluate eight LLMs from four model families evaluated on single-choice selection tasks. Beyond harmonizing heterogeneous tasks, AbductionBench separates final-answer evaluation from analysis of the observable reasoning traces and interaction histories produced alongside those answers. We analyze written traces as observable model outputs rather than faithful representations of internal computation, characterizing evidence use, hypothesis comparison, uncertainty, contradiction handling, and answer commitment. Jev matches or exceeds small LLMs when selecting among explanations over compact evidence but falls substantially behind on longer cases, suggesting that candidate discrimination and extended evidence integration place different demands on a system. In interactive tasks, the proportion of actions judged relevant declines at later steps, and a relevant first question is associated with a correct final answer; in the passive task, models rarely revise their initial hypothesis as new evidence arrives. Together, these findings highlight the value of evaluating abductive reasoning through both final answers and observable evidence use, hypothesis revision, and information gathering.
Evaluations of AI systems usually report properties of a model, while decisions to deploy concern a configured system: a model together with prompts, tools, permissions, data, safeguards and an operating context. We ask how evidence about technical failures, deployment-control effectiveness and real-world consequences can be connected to defensible authorization decisions throughout the lifecycle of a deployed system. We present R-BED (Risk-Based Evaluability Design), a specification-level methodology with a fixed 6+4+3 architecture of thirteen diagnostic dimensions (technical failure, deployment controls, consequence and impact); a failure-mediated expected-loss identity that is kept separate from intended-behavior and misuse pathways; bounded evidence statuses with independence levels; a control assurance ladder; five authorization dispositions governed by eight rules; and a Continuous System Evaluation (CSE) layer in which evidence carries a configuration pin, an expiry and declared dependencies, so that a change invalidates evidence in proportion rather than wholesale. We work through a synthetic refund-agent walkthrough, with a deterministic script that reproduces every number, to show how a task-success view can miss a consequential control weakness that the procedure identifies as missing evidence. The contribution is a reproducible methodological specification. Most components are standard practice integrated into one evidence-to-decision record; the Evidence Validity Register, the authority envelope and the tier rules are proposals. None has been empirically validated, and we state seven hypotheses for independent testing.
Neural networks trained by gradient-based methods learn the low-frequency content of a target function faster than its high-frequency content. This spectral bias is well documented, but it is much less clear which architectural choices mitigate it and which cannot. We study the question for shallow ReLU models linear in their trainable parameters, where each mode is learned at a rate given by a Gram-matrix eigenvalue, and for a general symmetric bilinear form $a(\cdot,\cdot)$, so that least-squares regression ($a=$ the $L^2$ inner product) and variational or physics-informed losses ($a=$ the $H^1$ energy form) are two instances of one analysis. Within this framework we propose gated local ReLU dictionaries, obtained by multiplying each ReLU feature by a hat partition-of-unity weight, in the spirit of partition-of-unity finite elements and overlapping Schwarz methods. The resulting Gram matrices are block-banded for any local bilinear form, with a block-Toeplitz interior whose matrix-valued symbol governs the spectrum. The two forms then behave very differently. For $L^2$ regression, the symbol is bounded away from zero, the ratio is uniform in the number of windows, and the measured spread of frequency-dependent learning rates falls from $1.9\times10^{5}$ to $4.5$ at matched model size. For the $H^1$ stiffness matrix, the symbol vanishes at zero frequency: gating leaves the $\Theta(N^2)$ conditioning intact and inverts the frequency ordering instead of removing the bias. A preconditioner built from the same interior stencil will fix the spectral bias, showing that architecture alone is not enough. Experiments on $1$D regression and Poisson problems confirm the dichotomy, and the regression effect persists when the hidden-layer parameters are trained as well.
While 3D Gaussian Splatting (3DGS) is typically optimized from images only, many robotics and mapping applications provide both camera and high-definition LiDAR data, combining rich visual appearance with accurate geometric structure. We introduce PLS-GS, a probabilistic LiDAR scaffold for Gaussian Splatting that exploits this complementary sensing setting to construct a compact 3DGS representation aligned to the measured surface. PLS-GS fits a Hierarchical Gaussian Mixture Model jointly over position and color, yielding a continuous and adaptive scaffold that eliminates the need for discrete scene partitioning such as voxel grids or octrees. The fitted mixture then guides the renderable 3DGS Gaussians: each rendering Gaussian is associated with a GMM component and constrained in the local surface frame induced by that component's covariance. A Mahalanobis support loss bounds the deviation of the rendering Gaussians from the corresponding LiDAR-derived surface. Across five large outdoor scenes, PLS-GS places primitives markedly closer to the LiDAR measurements than both image- and LiDAR-initialized baselines, while employing substantially fewer primitives, yielding a considerably more compact model and faster rendering. At matched capacity, PLS-GS places primitives closer to the LiDAR measurements, with comparable rendering quality. Code is available at https://github.com/k-harikumar/PLS-GS.
Selective unlearning is increasingly important for the safety and compliance of large language models. Harmful or memorized content is often entangled with useful knowledge and difficult to enumerate. We use the model's own failures to identify candidates for unlearning. We propose Hallucination-Guided Unlearning (HGU), which uses hallucinations as diagnostic signals. HGU localizes unsupported spans via evidence and internal consensus, traces a sparse set of high-responsibility MLP neurons, and edits them with compact gates under a utility-preserving KL constraint. On TOFU, HGU achieves the best MU-FE trade-off, with Avg scores of 0.7221 (forget05) and 0.7699 (forget10), and the highest SepS H-Avg (0.4382/0.4715) under mixed forget/retain prompts. On the WMDP auxiliary task, HGU yields the strongest forgetting (e.g., 14.2 on Economics) while improving Retain (52.7) and MMLU (57.2). These results suggest that tracing hallucinations enables targeted, post-hoc unlearning that better balances forgetting efficacy and general utility without full retraining.
Vision-language models pass text-based false-belief probes at near-human levels but drop by over twenty percentage points on abstract panels the moment the same problem is conveyed visually. The consensus diagnosis is conceptual, prescribing new belief-reasoning modules. We argue the dominant bottleneck is perceptual: not what the model cannot reason about, but what it cannot see. We propose a five-condition dissociation protocol that holds the reasoning task constant while varying perceptual fidelity from visual only to text only, and quantify the result with the Perceptual Contribution Ratio, the share of the visual-to-text gap a given fidelity closes. We test this on two kinds of scenes: abstract 2D panels in the Heider--Simmel tradition, and naturalistic 3D game scenes from MuMA-ToM and MindPower. The evaluation covers seven VLMs and roughly 159,000 API calls. Ground-truth scaffolds substantially close the gap in every setting, and on the abstract panels and MuMA-ToM they match or exceed the text-only reasoning. Whereas scaffolds built by automatic parsers, including frontier VLMs, fail. The most common error in the visual-only condition is \emph{reality bias}, which accounts for 60\% of false-belief failures. Models report where an object actually is, not where an absent agent believes it is. This points to a perceptual front-end that fails to track which agents were present at each moment. Under the evaluated protocol, structured visual scene-state extraction, rather than belief inference, is the dominant bottleneck for visual ToM.
We develop a variational and geometric theory based on statistical physics for the joint evolution of Transformer attention and token representations at fixed model parameters. To this end, we introduce a free energy for the attention and model the evolution of the attention weights by its Fisher--Rao gradient flow. At fixed token representations, we prove that the softmax matrix is the unique free-energy minimiser and that every positive attention trajectory converges to it. For prescribed time-dependent representations, we show that the attention dynamics becomes linear after a change to logit variables. This result yields tracking bounds relative to the instantaneous softmax and sufficient conditions for attention to converge to a fixed matrix. Hamiltonian dynamics for the representations and their momenta are then coupled to the attention relaxation driven by the same energy. This coupling results in an energy-dissipation law. We also develop two alternative Hamiltonians incorporating the attention value map. We prove global well-posedness of the attention flow and all three coupled systems.
Generating high-fidelity voxels while keeping their lightweight nature remains an open problem in computer vision. Current generators are either high-fidelity but computationally heavy, or lightweight but of lower fidelity. Hierarchical methods keep the lightweight nature of voxels. However, their coarse levels, which decide where all finer cells can appear, are generated with the least information about the shape and are never revisited, so they are often plausible but wrong. Our key insight is to let finer cells take part in deciding coarse cells. To this end, we propose JoVox, a discrete flow matching framework that learns the joint distribution of all levels of the voxel hierarchy, instead of a separate distribution for each level, so that coarse cells can be predicted with finer cells already generated nearby. Within this framework, we further introduce a two-level prediction, which outputs the cells two levels finer at once instead of one level finer, halving the number of levels that the network processes and yielding voxels of higher fidelity. We also introduce an unmasking strategy that revisits early decisions once finer cells are generated and corrects inconsistencies in masked diffusion models. Our framework outperforms the baselines on both unconditional and image-conditioned voxel generation. Compared with the current state-of-the-art TRELLIS.2, we achieve 8x higher training throughput, 1/15 of the FLOPs in a forward pass, train on 4 vs. 32 GPUs, and obtain higher quality with an 11% lower Chamfer distance.
Retinal screening programmes need to know which automated decisions can be accepted without review. This paper presents RetinaGuard, an agentic triage framework in which quality, diagnostic, uncertainty, verification and explanation agents gather evidence for each fundus photograph, a learned failure predictor estimates the probability that the decision is unsafe, auxiliary tools are invoked only when their expected value of information exceeds their cost, and the acceptance threshold is certified by conformal risk control. On the RFMiD test split, RetinaGuard was compared with confidence and calibrated-uncertainty referral at matched referral budgets and equal certified safety, using two RETFound primary models of different strength. With the weaker model (macro-AUROC 0.896), verifier disagreement added significant information, RetinaGuard captured 119 of 280 unsafe decisions at a 20% budget against 88 for lowest-confidence referral, and at a nominal 5% unsafe level it accepted 37.0% of cases against 10.3%. With the stronger model (0.925) the verifier added nothing beyond uncertainty and the advantage shrank to a few percentage points. In both settings its coverage at the nominal 5% level stayed near 37.7%, gating cut computation by about 30%, and a quality agent reduced unsafe acceptances under most image corruptions. The benefit of agentic evidence gathering thus depends on how much the primary model's own uncertainty already reveals about its errors.
Post-hoc merging of public LoRA adapters is an attractive gradient-free path to continual adaptation, but reliable measurement of catastrophic forgetting in this setting depends on infrastructure that prior work has not addressed. We treat each adapter merge as a sequential task in a continual-learning setting, with the instruction-tuned base model's capabilities as Task 0, and study forgetting on Llama-3.1-8B-Instruct. Public adapter hubs contain modules that fail basic provenance checks: silent base-version mismatches and near-zero delta norms whose inclusion inflates measured interference by ~10pp. We propose a two-check audit (base-model match, non-trivial delta norm) and a leakage-free probing protocol as preconditions for any forgetting study in this setting. On an audited 4-adapter set, naive merging causes -10.5pp on GSM8K and -18.3pp on HumanEval, while Task Arithmetic at lambda=0.5 sits within a standard error of the unmerged baseline. A coordinate search including negative coefficients identifies the math adapter as probe-optimal at lambda=-0.5, outside the reach of positive-only mergers, suggesting forgetting has structure that negation-aware merging may exploit. The same protocol replicates on Mistral-7B-Instruct-v0.3, where 2 of 4 candidate adapters fail the audit and the audited merge reproduces the same merge-method ranking on GSM8K, supporting that our findings are not Llama-specific.
Long-term personalized agents are increasingly expected to answer queries over user histories in which preferences, goals, and constraints evolve across interleaved conversations. Existing retrieval-based and structured-memory systems can surface relevant facts, but often return isolated memory units that do not explain which earlier preference was refined, contradicted, or superseded by later evidence. We introduce TransMem, a transition-aware memory retrieval framework that stores preference updates as explicit retrieval units alongside ordinary factual memories. During memory construction, TransMem routes user interaction history into topic-structured memory graphs, detects shift events, and grounds them as transition records. At inference time, TransMem first retrieves factual anchors and then adaptively expands over transition records, producing a compact context of factual anchors and accepted transition records. Experiments on PersonaMem and PersonaBench across two Qwen backbones show that TransMem outperforms representative retrieval-based and structured-memory baselines. Controlled analyses further show that transition records and adaptive expansion contribute beyond factual anchoring, chronological expansion, flat transition retrieval, and fixed-hop traversal. These results suggest that preference-update records are useful retrieval units for long-term personalization.
Document parsing benchmarks are often treated as stable measures of model quality, yet rankings may depend substantially on benchmark design choices. In this work, we present RankShift, a systematic study comparing OCR and vision-language models (VLMs) under controlled variation in dataset composition, matching strategy, and scoring methodology. Using OmniDocBench as a case study, we evaluate the ranking stability of 16 models across 3 dataset configurations, 6 matching strategies, and 3 scoring metrics. We quantify rank stability using pairwise rank correlation, top-k churn, and pairwise reversal rate. Our results show that the interaction between matching strategy and ground truth format is a substantial but understudied driver of instability: scoring against OmniDocBench's JSON annotations versus its Markdown rendering yields a mean Kendall's tau of approximately 0.33 under region-level matching strategies, with individual pairs as low as tau = 0.23, compared with tau = 0.77 when each page is scored as a single concatenated string. Dataset composition produces comparable instability across page-type strata (mean tau = 0.48) and layout-type strata (mean tau = 0.49), while the cross-dataset shift from OmniDocBench to Real5's Scanning reconstruction of the same pages is more stable (tau = 0.80). Metric choice is the most stable dimension, with mean tau of approximately 0.78 across NED, BLEU, and METEOR. These findings indicate that leaderboard position is not solely a property of model capability, motivating leaderboards that report not only headline scores but also the robustness of rankings across reasonable design choices.
Recently, diffusion models have achieved remarkable performance in image generation. However, their inherent memory properties introduce a potential safety risk of training data leakage, i.e., data extraction attack. To illustrate such risk, we propose a membership guided extraction (MGE) framework to effectively extract training data from a pre-trained diffusion model. Our core idea is to formulate the data extraction as a conditional generation, where the condition signal is from a membership inference classifier. To improve the success rate and efficiency of MGE, we propose adversarial gradient and parallel branch decoding. Our experiments on image generation benchmarks demonstrate that our MGE can extract hundreds of images, which shows that more attention should be paid to privacy-preserving of diffusion model.