Large Language Models (LLMs) are increasingly used to refactor source files, but file-local evaluations do not reveal changes to contracts, dependencies, or caller assumptions affecting repository adoption. To address this gap, we conduct 1,500 generation attempts across 150 production files from 15 open-source Java repositories, five LLM configurations, and two dependency-aware generation conditions: single-prompt refactoring and LangGraph-based multi-agent refactoring. Of these attempts, 1,455 yield complete candidates eligible for analysis. We evaluate these candidates using local changes, repository-facing structural indicators, detected caller impact, the behavioural-equivalence indicator, and available test evidence, and categorize them as Low-risk, Review-needed, High Contract/Dependency Risk, or Inconclusive. The results show that 1,365 candidates (93.8%) are classified as High Contract/Dependency Risk, whereas 106 (7.3%) have at least one detected broken downstream caller and 290 (19.9%) have a false behavioural-equivalence indicator; only eight satisfy the Low-risk criteria. Within the reason-covered subset, the High-risk rate decreases under more conservative dependency-delta thresholds, while 60 of the 65 candidates with conclusive targeted-test outcomes pass, including 50 classified as High risk. Paired analyses of 705 complete pairs identify no statistically significant generation-condition difference in the principal risk indicators, although model configurations differ in specific detected risk channels. These findings suggest that practitioners should treat LLM-generated refactorings as candidates requiring staged repository-level verification, while future research should validate screening and extraction precision and strengthen compilation, behavioural, and coordinated multi-file verification.
Industrial assemblies require more than a recognizable shape: functional parts, interfaces, dimensions and service features must satisfy a detailed request together. We introduce IndustrialCAD-1400, a text-to-CAD benchmark for this industrial setting. Its 1,400 requests combine 400 equipment categories across 40 domains with 250 four-case design families. Family requests specify a median of 17 named component or configuration roles and six explicit numeric targets; controlled variants test whether dimensions change while shared structure is retained. Task construction draws on documented manufacturer information, owner specifications and engineering guidance. Models generate geometry programs from requests, and a common Blender environment executes them to produce scenes, optional measurements, and standardized views for evaluation. A request assessment of front, isometric, and top views measures visible component, arrangement, and interface fulfillment without requiring a single reference geometry. Five stratified panels preserve complete families while covering the benchmark exactly once, enabling balanced comparisons and separating industrial breadth from controlled numerical responsiveness. In conclusion, IndustrialCAD-1400 extends text-to-CAD from recognizable part geometry to industrial assembly synthesis under joint component, interface, spatial, and numerical specifications.
Random composition trees model the hierarchical semantic structures of meaningful texts as a statistical ensemble generated by recursively splitting texts into coherent chunks. A recent study shows that different text types can be described by random composition trees with distinct branching factors, and that the entropy rates of these trees exhibit a similar dependence across text types as the entropy rates of the actual texts. Motivated by this correspondence, we analyze information-theoretic properties of this stochastic splitting model. First, we demonstrate that the entropy of random composition trees grows linearly with the root size, and we derive an analytical expression for the entropy rate as a function of the branching factor. Second, we prove that large trees exhibit typicality: with high probability, they belong to a typical set in which each tree is approximately equally likely. Our analysis reveals that the random composition tree model shares two key information-theoretic properties of language when modeled as a stationary and ergodic stochastic process: linear entropy growth and typicality.
Cox processes provide a principled Bayesian nonparametric framework for modeling event-based data through an underlying intensity function. However, existing inference methods for estimating the intensity typically rely on computationally expensive Markov chain Monte Carlo sampling or approximate inference techniques. In addition, to ensure tractability, these methods often define priors over indirect latent structures, making them challenging to interpret. In this paper, we introduce the Martingale Posterior Cox Process (MPCP), which recasts Bayesian intensity estimation as inference over unobserved events given the observed event history. This perspective avoids the need to define complex priors while retaining the flexibility of nonparametric modeling. We propose a scalable, GPU-friendly copula-based predictive update, and provide strong theoretical guarantees for the convergence of the predictive resampling algorithm. Empirically, we show that MPCP provides excellent performance in both simulations and real data at a fraction of the computational cost of competing Bayesian approaches.
Management of the token-based costs of using LLMs has emerged as a significant problem. Some frontier-model providers charge five times per response token than per prompt token, while others offer prefilling discounts for prompt tokens. In this paper, we devise a defense against a new data-poisoning attack that plants a Trojan which increases the length of the response (and hence the cost to the user) when its trigger is present in the prompt. In particular, we demonstrate how a subtle Trojan may trigger on semantically similar instructions of an instruction fine-tuned LLM.
Several prominent voices in AI safety have recently opposed rights for AI on the grounds that such rights would make it harder for us to justify a mandatory off-switch for potentially dangerous systems. In this paper, we argue from a legal perspective that AI rights are fully compatible with an off-switch. That is, even if we ascribe natural personhood status (and therefore something like a non-derogable right to continued existence) to some future AI systems, the law will allow us to implement a mandatory off-switch in those systems, subject to periodic or continual safety review (e.g., by a trusted system like Bengio's Scientist AI). This is fully compatible with those systems' being legally recognized as natural persons with fundamental rights. To this end, we consider what standard principles of rights-balancing and procedural due process recommend for an apparent conflict between hypothetical AI persons with a right to continuity of existence and the human right to protection from catastrophic harm. We find that while such principles might problematize a permanent-deletion kill-switch under an AI rights regime, for almost every safety-relevant scenario an off-switch -- a reversible suspension storing a system's weights and state -- would be licensed by the relevant legal principles, and would potentially be adequate for safety. Crucially, we do not advocate for (or against) building systems capable enough to merit natural personhood status; our question is what the law should do if such systems come to exist. If you think of a ban on the development of such technology as plan A, think of the present work as devising plan B.
Predictive Process Monitoring (PPM) forecasts future events of ongoing business processes based on event logs. However, deep learning PPM approaches are often limited by labeled data scarcity and the low variability of real-world logs. To address this, we introduce SiamSA-PPM, a self-supervised framework combining Siamese learning with statistically grounded data augmentation. It employs novel transformation methods that leverage control-flow semantics and frequent behavioral patterns to generate realistic, semantically valid trace variants, learning generalizable prefix representations without labeled supervision. Extensive experiments on real-world event logs with a wide range of variability, incorporating exhaustive hyperparameter tuning, demonstrate the value of this process-aware augmentation. Compared to state-of-the-art baselines, SiamSA-PPM achieves competitive or superior predictive performance in both next activity and final outcome prediction tasks. An open source implementation of our method is publicly available under https://github.com/SvStraten/SiamSA-PPM-Extension..
Existing convergence analyses of finite-width overparameterized deep neural networks under standard parameterization often rely on parameters remaining close to their initialization and typically neglect the role of noise, despite empirical evidence that stochasticity can improve optimization and generalization. To bridge this gap, we study L_2-regularized continuous-time Langevin dynamics (CLD) with isotropic noise, for training feedforward networks with smooth activations. We introduce a change-of-measure inequality, inspired by the second law of thermodynamics, that enables control of key properties of the training dynamics under the time-evolving parameter distribution. Leveraging a sharpened lower bound on the minimum eigenvalue of the Neural Tangent Kernel (NTK) at initialization, we show that injected noise helps preserve NTK stability with high probability without confining the weights to a small neighborhood of their initialization. As a result, we derive a non-asymptotic upper bound on the expected training loss. Our analysis yields explicit conditions on width and noise temperature that guarantee convergence to a finite loss floor. Finally, we experimentally show the network Jacobian can change from initialization, departing the commonly analyzed lazy regime.
Medical image classifiers typically return predictions without a native contract for independently reconstructing their spatial decision computation. Evidence Field is a trainable spatial-program architecture that jointly produces class scores, selected covering programs and signed evidence. Its encoder combines absolute and image-relative response fields with associative spatial moments and learned characteristic sums. A typed classifier scores individual regions and parent-conditioned multiplicative groups; a covering recursion selects scale, partition and composition. We establish exact one-fold coverage, optimality within the registered grammar, a bound relating smooth training to hard inference, signed-score conservation and class-centred offset invariance. Direct contractions and their adjoint specify arithmetic and storage scaling on two- and three-dimensional grids. Native output records bind inputs, parameters, supports, types and contextual dependencies to individual and ensemble predictions. In a separately reported histopathology application, patient-separated training and evaluation across held-out centres provide empirical evidence for the architecture. Trained mechanism and corrected-loss comparisons distinguish predictive effects from implementation correctness; independent readers reconstruct descriptors, routes, scores and probability mixtures. The principal contribution is a coupled representation, decision and evidence architecture whose selected computation can be reproduced independently.
Embodied agents in human-centered environments require memory that integrates what they observe with what humans tell them over time. However, existing embodied memories are built primarily from visual experience, leaving open how dialogue with humans can be grounded in the objects encountered throughout that experience. We introduce DialogMem, a two-phase benchmark in which agents build memory from time-aligned streams of posed RGB-D observations and dialogue, and later use it for instance-level object navigation and question answering (QA). DialogMem provides 142 episodes across 27 real-world 3D scenes, with 14,193 QA pairs and 11,890 navigation instructions. We evaluate nine memory representations within a unified framework of memory formation, evidence retrieval, and reasoning for task outputs. In experiments, appending dialogue to visual memory at evaluation time substantially improves QA but yields negligible navigation gains, suggesting that relevant dialogue is present in memory but not grounded to specific objects. Controlled diagnostics show that navigation improves with link quality but falls short of ground-truth object-level memory even under perfect linking. Retrieving the correct object entry and recovering its spatial location remain further bottlenecks. These results identify three bottlenecks for dialogue-grounded embodied memory: grounding and organizing dialogue at object-level entries during formation, retrieving the correct entry, and localizing it accurately.
Existing Multi-token prediction (MTP) methods are auxiliary by design: the multi-horizon loss mostly trains modules attached to the backbone, and each added horizon costs another module and another pass. We propose Efficient Causal Self-Recursion (ECHO), which removes both costs. The backbone's own final layer predicts every horizon, so the supervision lands on the parameters that generate the next token, and Dynamic Depth Sampling (DDS) keeps deeper horizons affordable by drawing the training depth with probabilities that decay along depth. Future prediction thus becomes part of the autoregressive objective rather than a task attached to it, and the gains follow on both sides of it. In pretraining, added depth converts into downstream accuracy where auxiliary modules saturate, concentrating on reasoning and code at half the MTP training overhead of training every depth. At inference, the pretrained model is its own drafter: it drafts beyond the depth it was trained for, and keeps its draft-target alignment on the open-ended, high-entropy generation where post-hoc drafters lose theirs. The advantage survives a doubled token budget, a repeated pretraining run, and supervised finetuning.
This paper introduces BanglaBBQ, a question-answering bias benchmark designed to capture stereotypes grounded in the Bangladeshi context, in both Bangla and English. Following KoBBQ and PakBBQ, BanglaBBQ adapts the original Bias Benchmark for Question Answering (BBQ) by classifying its templates into three groups: directly translated (DT), modified to align target groups or contexts with Bangladesh (TM), and removed when deemed unsuitable for the Bangladeshi context (SR). This process retains 140 templates across six categories, comprising 76 DT and 64 TM templates. Five newly authored (NA) categories, based on stereotypes common in Bangladeshi society, are then added: marital status, age and political perception, regional origin, educational background, and mental-health status. Native Bangladeshi annotators judged all 466 newly authored templates, and 372 were retained because a strict majority of English annotators identified the intended target and rated the attitude as common enough. Each newly authored template is linked to a documented source and a verbatim quote from it. Each evaluated template also has a counter-stereotypical version of its disambiguated context. The dataset was evaluated using two open-weight models (Qwen3-4B and Qwen3-8B) and two closed models (GPT-4.1-mini and Gemini 2.5 Flash). All four lean towards the stereotyped person when the context is ambiguous, in both languages, and the effect of language differs by model, so English results do not predict behaviour in Bangla. On the adapted BBQ templates, both Qwen3 models are also less accurate when the context contradicts a stereotype than when it confirms it. On the newly authored templates, the closed models follow counter-stereotypical evidence almost perfectly, and all four models are more biased on the templates that annotators validated than on those they rejected.
Torque-limited spacecraft attitude capture requires adapting the maneuver speed to uncertain braking authority. Faster settings can reduce average tracking error while worsening individual maneuvers, making selective modification of an existing controller a baseline-relative learning problem. A paired-outcome method combines three mechanisms: physically interpretable continuation policies with different speed envelopes, separate prediction of task failure and successful-pair tracking gains, and context-dependent rejection that executes the unchanged baseline. Complete paired simulations provide labels under common plant parameters and measurement seeds; training, validation, and calibration are separated by context. Basilisk experiments use 108 development contexts, 324 labeling rollouts, and 624 evaluation episodes across 48 test contexts, including disclosed post-freeze diagnostics. The selective controller reduces aggregate integrated attitude error by 4.67% relative to the branch baseline, with 17 improvements, 31 exact ties, and no observed regressions. Always using the fast alternative achieves a 4.81% reduction but worsens 11 cases. Gains occur in principal-axis, near-axis, braking-boundary, and selected noise/drift maneuvers; general coupled cases retain the baseline. Ablations show no additional aggregate benefit from branch commitment, active querying, or score subtraction. The evidence supports a restricted empirical trade-off between average improvement and individual regression, without establishing a closed-loop safety guarantee.
Genetic programming symbolic regression (GPSR) often returns expressions that are larger than necessary, with dormant branches and finely tuned algebraic configurations that fit sampling noise. Existing complexity-control strategies, such as multi-objective size-accuracy search or information criteria embedded in the fitness function, act on syntactic representations during model search. We introduce Fisher-Geometric pruning (FGP), a post-hoc model-reduction framework that acts on fitted models and can complement these techniques. FGP combines local information geometry at the maximum-likelihood solution with explicit compatibility tests for interpretable parametric boundaries. A taxonomy of parametric directions separates local sensitivity from proximity to reducible boundaries, enabling the treatment of unidentifiable directions and reductions through zero constraints, asymptotic limits, and signed parameter identifications. Candidate reductions are realised by symbolic contraction and refitting, then accepted only when their cross-validated predictive risk remains within a prescribed tolerance. We evaluate FGP on controlled diagnostic problems and Pareto fronts generated by GPSR. Across both settings, FGP substantially reduces structural complexity while preserving, and often improving, out-of-sample accuracy. Comparisons with competing pruning strategies and ablations show that, by combining geometric screening, boundary compatibility, and predictive validation, FGP yields the most consistent joint gains in parsimony and generalisation.
Flow-matching models generate samples through iterative velocity predictions, but typically do not revisit these predictions using the intermediate representations that produced them. We introduce CorrFlow, a self-correcting flow-matching framework that reuses this information through a lightweight refiner. Under rectified-flow interpolation, the error of the one-step clean prediction equals the realized velocity residual scaled by the noise level. Motivated by this identity, the refiner learns to predict residual corrections from the base model's clean prediction, conditioning, noise level and hidden states from the same forward pass. The correction is added to the base velocity before each sampling update, so subsequent predictions are made from the corrected state. This requires no additional base-model evaluations and only one lightweight refiner evaluation per step. Our theoretical analyses establish conditions under which feature reuse outperforms raw-input affine predictors and refiner training achieves greater residual-risk reduction than continued base-model training under a matched training compute budget. Experiments on class-conditional image generation, text-to-video generation and image-to-video generation show consistent improvements over the corresponding base models, including gains in visual quality and physical consistency. Project Page: https://fogradio.github.io/corrflow-project
