Papers

Language models are typically pretrained from random initialization. Recent work challenges this convention, showing that a brief warm-up on abstract, algorithmically generated data can provide a better starting point for subsequent learning of natural language. In this paper, we show that in small language models, such a warm-up improves specific capabilities that are not reflected in language-modeling perplexity. Our warm-up uses an abstract stack-manipulation task that requires compositional and state-tracking capabilities. Allocating as little as 1% of pretraining tokens to this data improves multi-hop question answering by up to 3.9 F1 points on MUSIQUE, with additional gains on HOTPOTQA and 2WIKIMULTIHOPQA despite comparable language-modeling perplexity. Controlled experiments show that the warm-up substantially accelerates the acquisition of deeper reasoning chains. We also explore what drives this transfer. First, the structure of the data matters: replacing the stack task with a queue fails to produce the same gains. Second, the gains are specific: performance improves on sequential reasoning chains, with no consistent benefit on tasks that combine or compare independent facts. Third, timing matters: mixing abstract data with natural language is far less effective than an initial dedicated phase, and exposure after pretraining completely removes the benefits. The early advantage persists through billions of subsequent language tokens. These results show that early abstract training can reliably shape the capabilities language models later acquire.

Predicting molecular properties for compounds that differ structurally from labeled training molecules is important for drug discovery and materials design. Tabular foundation models (TFMs) offer a promising approach through in-context learning, but their performance under structural shifts and the value of molecular comparisons in this setting remain underexplored. We study structural generalization in molecular property prediction and introduce MolPAIR (Molecular Pair-Augmented In-context Refinement), a framework that combines molecule-level and molecular-pair contexts without task-specific parameter updates. A global tabular foundation model (TFM) first predicts a query's property from labeled molecular examples. A second frozen TFM predicts differences in prediction errors between the query and labeled reference molecules, using these comparisons to refine the initial prediction. Across 58 MoleculeACE and Polaris tasks, CheMeleon representations combined with TabPFN-3 already outperform each evaluated baseline on a majority of tasks. MOLPAIR further improves this predictor on 46 of 58 tasks, with gains across four molecular representations and three TFM backbones. These results show that explicit molecular comparisons can strengthen tabular in-context learning for structural generalization while keeping the molecular encoder and pretrained model weights fixed. The code and datasets are available at https://github.com/nums-ai/MolPAIR.

Recent LLM-based approaches to control either invoke a language model to select actions or synthesize world models that require planning at every decision, introducing latency that can limit real-time use. We introduce Code to Control, an approach that synthesizes Python controllers which execute directly as policies. Code to Control separates program structure from parameters. An LLM synthesizes the controller structure, while derivative-free search fits its parameters for continuous control using feedback from the environment. Once learned, the resulting controllers require neither LLM inference nor planning at decision time, enabling real-time gameplay and, under our timing protocol, faster action selection than a PPO policy. Across a suite of Atari games, Flappy Bird, and MuJoCo tasks, Code to Control outperforms planning-based program synthesis methods, remains competitive with deep reinforcement learning while using fewer environment interactions, transfers across substantial changes in environment dynamics, and scales to complex locomotion tasks.

Text-to-image diffusion models are personalized to a subject by DreamBooth fine-tuning on a handful of its images. Increasingly, these images come from a diffusion model rather than a camera. We show that fine-tuning on such synthetic images degrades subject fidelity, producing oversaturated color and excess high-frequency detail. To isolate the cause, we fine-tune two models from the same base model with the same DreamBooth recipe, one on real photos of a subject and one on synthetic images of that subject generated by the first. We trace the degradation to classifier-free guidance (CFG). For the model personalized on synthetic images, the angle between the conditional and unconditional noise predictions, and with it the norm of their difference, is much larger than for the model personalized on real photos. This inflation grows toward high frequencies and also appears at other prompts semantically close to the subject, such as its class noun, but not at unrelated ones. We propose ReGain, a training-free correction applied at sampling time that measures how much each frequency band of the guidance is inflated relative to the base model and scales that band down accordingly. ReGain needs no real photos. On Stable Diffusion v1.5, ReGain closes 51-64% of the subject-fidelity gap to the model personalized on real photos, as measured by DINO, DINOv2 and CLIP-I. It also improves subject fidelity on SDXL and SD 3.5 and preserves text alignment on all three backbones.

Weakly supervised localization on chest radiographs is reported as a single overlap accuracy, and those numbers are not comparable across papers. We measure what one contains. Against position-null controls that keep a readout's own box and move only its position, class activation mapping and its gradient-weighted variant beat the null on two to four of seven annotated findings on ChestX-ray14 and two of eight on CheXlocalize. The null is beatable: three board-certified radiologists clear it on four to seven. Three rarely stated choices each move a result by more than most reported method effects: the results regime, the extraction threshold, and the mask-to-box rule, whose cost is predicted by how fragmented the thresholded mask is. Predicted box area alone reverses a paired comparison. These measurements produced PACE, for population anomaly and class evidence, which multiplies the finding identity from the label-trained head by the spatial evidence from a frozen radiograph backbone that never sees a pathology label. It is the case study that made the protocol dependence visible, and we report its costs beside its gain: in an exploratory matched-area analysis it improves strict-overlap accuracy on both corpora and beats a label-derived second map, but it is the least faithful of the three readouts to the classifier's causal map, its advantage depends on the box rule, and, this benchmark releasing no localization validation set, every design choice behind it was made on the boxes it is scored on. Its one clean transfer beats one baseline and loses to the other.

Query-based segmentation models achieve strong performance even under real-time inference on workstation GPUs. On edge devices, however, we show their transformer decoders are deeply memory-bandwidth-bound, a bottleneck invisible to standard FLOP analysis that worsens with input resolution and limited DRAM bandwidth. An FP16 roofline analysis with tensor-core GEMMs active and hardware DRAM counters shows that the non-tensor-core aggregates of FastInst and Mask2Former reach only $2.5\%$ to $6\%$ of the CUDA-core ridge yet account for $53\%$ to $70\%$ of decoder time and $61\%$ to $71\%$ of decoder bytes. The largest individual kernels sustain up to $90\%$ of nominal peak DRAM bandwidth, effectively saturating the memory bus. The bottleneck is therefore not primarily low GEMM intensity, but cross-kernel activation traffic. CompactInst addresses this bottleneck using parameter-free spatial pooling at $4{\times}$, which reduces the bytes moved within the transformer decoder, while a two-convolution refinement head recovers spatial detail. Under TensorRT FP16 on Jetson Orin Nano, CompactInst delivers decoder speedups from $2.35{\times}$ to $4.74{\times}$ and model speedups from $1.57{\times}$ to $2.62{\times}$ as input resolution grows, while closely matching accuracy ($-0.4$ AP on COCO val2017, $+0.2$ AP on Cityscapes test). A heavier variant with $2.57{\times}$ more decoder FLOPs still runs its decoder $1.74{\times}$ faster on the most bandwidth-constrained device, confirming at the deployment level that decoder bytes, not FLOPs, govern latency.

The development of Financial Natural Language Processing (NLP) systems for Brazilian Portuguese is hindered by the scarcity of domain-specific datasets. To address this, we present a large-scale corpus of approximately 160,000 Brazilian Portuguese financial news articles collected from web sources. To establish initial benchmarks for a weakly supervised multi-class classification approach based on semantic similarity, we annotated subsets of this corpus both manually and using a Large Language Model (LLM). We evaluated three similarity-based methods – cosine similarity, K-Nearest Neighbors (KNN), and FAISS-based nearest-neighbor search – using multilingual semantic embeddings. Our experiments reveal that the task is highly sensitive to label quality and class distribution. On the manually labeled subset, KNN achieved the best performance (47.3% accuracy, 31.9% F1-score), closely followed by FAISS. Conversely, on the LLM-annotated subset, while accuracy reached approximately 50%, F1-scores dropped to near 20%, highlighting the negative impact of label imbalance and noisy weak annotations. These findings establish foundational baselines for a challenging task, underscoring the critical need for high-quality manual annotations. Ultimately, this dataset and our preliminary results provide a reference point for future research in Brazilian Portuguese Financial NLP.

Activation steering controls large language model behavior by adding a direction to hidden activations during inference. For steering to work well, that direction must be a pure signal for the target concept: any confound entangled with it, such as an incidental topic or writing style, is injected into the residual stream alongside the concept itself and degrades the generated text. We introduce Contrastive Covariance Steering (CCS), a training-free method that eliminates this confound variation. CCS projects steering vectors to a low-dimensional subspace obtained in closed form by solving a contrastive variance objective, which retains the directions along which the target concept varies while suppressing directions shared with the other classes. Across three LLMs and four steering domains, CCS consistently beats all baseline steering methods in all $12$ model-domain settings. The largest relative improvement occurs on AxBench and Emotion where we outperform the next strongest steering baseline by $48$ to $70\%$ and $13$ to $27\%$, respectively. Furthermore, CCS outperforms naive difference-of-means steering by up to $151\%$, showcasing the importance of deconfounding concepts. A toy experiment verifies our motivation on why: planting a topic nuisance into the training data reduces unprojected steering by $46\%$, whereas CCS loses only $12\%$ and recovers most of the gap to an oracle that explicitly deletes the planted direction. A human study with $23$ raters on $296$ generations reproduces the LLM judge's ranking. Our extensive ablation verifies the importance of our method's components and robustness to the number of nuisance classes. Finally, we extend our method to multi-turn conversations, where CCS sustains both concept expression and fluency, matching system prompt performance as other steering methods drift.

Physics time-series data are central to fundamental discoveries involving dark matter, neutrinos, and gravitational waves. Machine learning has shown strong potential to recover weak scientific signals buried in complex noise. Recent LLM-agent systems further automate ML development, but existing approaches are often driven primarily by one or a few numerical scores. In pursuit of a better score, agents may exploit shortcuts, trigger mode collapse, or game the evaluation procedure, producing models that achieve high scores but are scientifically invalid. We introduce SIDERIUS, an LLM agent system designed to move beyond a better score toward scientifically valid, long-horizon ML development. SIDERIUS combines contract-defined composable capabilities for flexible orchestration and human-in-the-loop, a multi-layer scientific evaluation protocol that distinguishes valid models from unhealthy or gaming models, and resource-aware exploration for computationally expensive scientific ML. In our TIDMAD benchmark, SIDERIUS with a GPT-based orchestrator achieves the highest Valid-model proposal rate of 63.8% and the best valid denoising score of 2.744, compared with 1.6% and -0.647, respectively, for the same GPT model using the general-purpose Codex coding agent. We further benchmark SIDERIUS on three additional physics time-series datasets, demonstrating its reuse and performance across different tasks. Together, these results suggest that progress in agentic scientific ML should be measured beyond a better score, by how an agent explores and by whether the models it produces can be trusted as science.

Predicting which scientific connections will form next can point researchers toward the most promising leads. One successful approach learns representations by sampling paths through publication networks, represented by a hypergraph, that connect scientific entities, concepts, and researchers, approximating how knowledge and expertise move through the literature. Sampled paths are then used to train an embedding model to forecast scientific connections such as a property of a material. Performance, however, depends strongly on how many paths are sampled: too little sampling yields incomplete representations, while further sampling can eventually weaken prediction quality. The best sampling budget is therefore often identifiable only after future discoveries are known. We introduce Query–Anchor Agreement (QAA), a diagnostic requiring no future labels or separately trained model. QAA can be used to select a sampling budget, and refine or re-rank the resulting forecast to improve forecasting scientific connections. QAA compares the candidates a model predicts for a target with those most similar to the target's known historical examples, and identifies when these two views converge. Across thermoelectric-material and COVID-19 drug-repurposing benchmarks, repeated random-seed runs, and three sampling policies, QAA-selected budgets fall on average 1.86 percentage points below hindsight-optimal selection for broad candidate recovery. Further, promoting candidates supported by both views, under a capped fusion that bounds the anchor contribution, improves broad recovery across all six benchmark-policy settings, which define the types of network walks used for sampling.

On-policy distillation (OPD) improves language-model by supervising a student on prefixes sampled from its own policy, but its supervision remains confined to the output space. Consequently, OPD transfers what a teacher predicts without supervising the intermediate computation that produces those predictions. Simply distilling intermediate representations like representation matching is insufficient since teacher and student representations may differ geometrically, and even well-aligned hidden states need not be usable by the student's subsequent computation. We introduce On-Policy Mechanism Distillation (OPMD), which transfers teacher-derived intermediate computation while explicitly accounting for its compatibility with the student. OPMD first constructs a student-compatible teacher that transforms teacher representations into intermediate states usable by the student's downstream computation, then it internalizes these states from on-policy trajectories. OPMD achieves the highest average accuracy among the evaluated distillation baselines in all three teacher–student settings, and generalizes to multi-teacher distillation and cross-family supervised fine-tuning. Further analyses show that mechanism supervision yields more effective reasoning and meaningfully alters the student's intermediate computation and solution strategies, supporting that it transfers information beyond output supervision alone. Together, these results establish mechanism-level supervision as an effective way to move on-policy distillation beyond output supervision, enabling students to learn inner mechanisms they can use and internalize.

Training multi-agent collaboration requires trajectories of checking, feedback, and repair, yet task-solving rollouts may end in immediate approval or repeat unsuccessful attempts without useful correction. We introduce CoEvoke, a cross-domain framework that evokes these interactions during trajectory collection. Controlled perturbations create recoverable difficulties, while restricted reference-based guidance directs investigation. The resulting dataset contains 4,567 reconstructable trajectories over 4,130 software-engineering and reasoning tasks, linking public interactions to evidence, decisions, and task-native outcomes. Separate experiments over 2,287 same-task pairs show feedback increases of 47.72-100 percentage points across seven perturbation groups; guidance has smaller, model-dependent effects. With identical training tasks, decision-record counts, and updates, perturbation-enriched judge training improves ACPBench-Hard success from 35.38% to 42.12% (+6.73 points). SWE verifier training gains 7.60-9.40 points over interface-matched untrained references. Workflow and post-training evaluations preserve task-solving topology with unperturbed inputs and no private-reference guidance. These results support designing collection conditions to obtain corrective supervision that improves collaborative decisions with frozen solving agents.

We propose Barycentric Anchored Search, a post-hoc weight-space procedure for improving fine-tuned vision models under distribution shift without additional gradient training or increased inference cost. Given a pre-trained initialization and two checkpoints fine-tuned on a source dataset, the method constructs barycentric candidate weights and performs a small derivative-free search along directions connecting each candidate to the pre-trained model. Candidate weights are selected using a small validation subset of one or more out-of-distribution (OOD) datasets; the selected model is then evaluated on the held-out portions of those datasets, on additional OOD benchmarks, and on the ImageNet validation set to quantify the robustness-accuracy trade-off. The method is inspired by simplex-based derivative-free optimization, including Nelder-Mead and Torczon multidirectional search, but is specialized to the geometry of neural network weight merging. Unlike Model Soup, which typically benefits from many fine-tuned checkpoints, and unlike Model Stock, which applies a fixed geometric merge, our approach uses only two fine-tuned checkpoints and iteratively adapts the merge using validation feedback from OOD data. Experiments with CLIP ViT-B/32 show that the proposed search substantially improves performance on held-out OOD data, including ImageNet-A, ImageNet-R, and ImageNet-Sketch, while keeping ImageNet accuracy competitive. The resulting weights also improve transfer to OOD datasets not used during candidate selection, outperforming the original fine-tuned checkpoints, naive model soups, and Model Stock in our evaluations. These results suggest that lightweight optimization in anchored weight space provides a practical alternative to fixed weight-averaging rules, geometric merging, or single-model selection.

Most existing sports datasets are individual-centric or ball-centric, focusing on individual player actions while ignoring interactions among teammates. In badminton doubles, tactical success is governed by rapid team formation and coordination, yet standard resources for modeling these player interactions are scarce. To bridge this gap, we present DShuttleNet, a comprehensive dataset for badminton doubles curated from elite BWF World Tour tournaments. Beyond trajectory data, DShuttleNet uniquely incorporates expert-labeled "Loss Reasons," transforming the dataset into a diagnostic tool for causal tactical evaluation. We formally define two geometry-driven representations, Teammate-Relative Angle (TRA) and Vertical Shot Angle (VSA), to characterize coordination and shot aggressiveness. To quantify interpersonal synergy, we introduce novel metrics, Court Coverage Efficacy Index (CCEI) and Threat Load Asymmetry Index (TLAI), which evaluate the often-overlooked non-hitting partner's role and defensive coordination. Experimental results demonstrate that conventional models suffer from underfitting when failing to incorporate these coordination-centric features, highlighting the critical importance and potential impact of releasing DShuttleNet. Dataset is available at: https://doi.org/10.34740/kaggle/dsv/16097938.

Learning-to-rank (LTR) models can align rankings with undesignated signals when those signals appear as features or leak into training labels. We compare label-free genetic multi-criteria sorting (open Pikaia strategy variants), which reranks fixed candidate pools using an explicit criteria vector C, to label-trained LTR (LightGBM, XGBoost, RankSVM) under two controlled protocols on MSLR-WEB10K, Baidu-ULTR, WANDS, and Expedia. In the main proxy-skew study, labels remain clean while a calibrated proxy P is available as a feature: Genetic AI rankings show lower Spearman rank-proxy correlation than the LTR peers (e.g. about 0.15-0.16 vs. 0.59-0.87 on WANDS), with both families evaluated jointly on clean nDCG@10 on a bias-utility frontier. In a supporting label-contamination study, training labels are mixed with contaminant B: because Genetic AI methods do not use labels, their bias and utility curves stay flat, whereas LightGBM and XGBoost increase rank-B correlation as contamination rises (OLS bias_slope about 0.58-0.65 on WANDS/Baidu). The results support genetic sorting as an architectural option for optimizing declared attributes on a fixed pool when undesignated proxies or label contaminants must be refused.

Accurate day-ahead photovoltaic (PV) power forecasting requires both strong temporal modeling and adherence to physical PV system constraints. Purely data-driven models often ignore the underlying PV conversion process and may produce predictions that are outside the operational range of a PV system. In this paper, we propose a physics-informed state space framework that combines a differentiable PV conversion model with a selective state space backbone (PowerMamba). Our approach follows a physics-residual decomposition. The physics module provides a structured baseline based on solar geometry, irradiance, temperature effects, and inverter limits, while the state space model learns residual corrections from numerical weather prediction (NWP) inputs over a 24-hour horizon. Experiments on one year of utility-scale PV data show that the proposed method significantly improves forecasting accuracy, reducing RMSE by 38.5% and MAE by 45.8% compared to a physics-only baseline, and outperforming a purely data-driven model. Results further show that incorporating system-level physics provides the largest performance gains, while additional refinements offer only marginal improvements. Beyond accuracy, the model improves robustness by constraining predictions toward physical feasibility, reducing unrealistic predictions under corrupted measurements.

Drug development is an expensive and lengthy process requiring developers to select drugs for human trials and diseases to target. Computational tools that can estimate before testing which drug-disease candidates are most likely to reach FDA approval could reduce development time and costs through early program advancement and disease prioritization. Yet existing approaches predict outcomes at a single level or stage or do not distinguish between disease outcomes for the same drug. We ask (1) whether FDA approval can be forecast for a drug-disease pair using only pre-clinical trial information, and (2) whether these forecasts can help prioritize diseases for the same drug. We built a cohort of 14,134 drug-disease pairs by reconstructing programs from ClinicalTrials.gov, extracting disease-level approvals from FDA label documentation, and enriching each pair with molecular, disease, ADMET, target, and pathway features. We divided the cohort into pre- and post-2019 programs and trained a classifier on pre-2019 data to predict FDA approval. On the post-2019 cohort (2,637 programs), the model would have prioritized a program that ultimately gained approval over one that did not approximately three out of four times. For drugs tested across multiple diseases, selecting the model's highest-ranked disease would have identified an ultimately approved disease 61.3% of the time, compared with 40.7% under random selection. These results show that public pre-trial information contains signal for forecasting FDA approval at the drug-disease level and can help prioritize diseases for the same drug. We provide a public web tool for SMILES- and ICD-10-based drug-disease approval forecasting.

Many systems read the latent space of a variational autoencoder (VAE) as discrete codes, and rely on each latent being read back as the code it was written as. Shannon (1948) took such a shared codebook as given; inside a VAE nothing enforces it. We prove that no item of a standard training report can reveal whether the written and read codes agree on encoded data. We construct two VAEs that reconstruct every input perfectly, infer exactly, and match on every per-example ELBO and every code histogram, yet agreement is one in the first and zero in the second. We then show that the ELBO's own terms bound how far a measured agreement can move: the variational gap bounds the shift to exact inference, and the rate bounds the shift to generation from the prior. For Gaussian posteriors, a mean-field encoder pays an exact price in the gap that depends on its choice of axes. On probabilistic PCA the gap certifies a trained encoder in all 75 settings we test, at agreement between 0.76 and 0.97. On 240 trained VAEs, a floor set by the rate and the dataset size rules out rate-side certification in advance for 66 models, and none of them certifies. The results yield a simple protocol, Certify-or-Measure, for any pipeline that reads a VAE as codes.

Modern experiments increasingly produce rich post-treatment objects—such as text, images, and video—that may provide early surrogate signals for outcomes that are delayed, expensive, or observed only for a subset of participants. A pretrained encoder can turn such an object into a strong generic embedding, but the embedding is not trained for the estimation tasks required by a causal estimator and may emphasize information irrelevant to the causal target. We propose SIRNET, a Surrogate-Informed Representation adaptation framework for neural-NETwork-based causal estimation in the presence of unstructured post-treatment surrogates. Starting from pretrained surrogate embeddings, SIRNET learns a shared adapter jointly for outcome prediction, treatment propensity, and arm-specific surrogate marginalization, beyond the generic pretraining objective. The resulting covariate and surrogate representations provide a common interface for these nuisance tasks. The final treatment-effect estimate and confidence interval are obtained from cross-fitted orthogonal scores that attain the semiparametric efficiency bound under the stated nuisance-rate conditions. We also validate SIRNET's performance in synthetic and semisynthetic experiments, as well as the Anti-Amyloid in Asymptomatic Alzheimer's Disease (A4) trial, using frozen pretrained brain-MRI embeddings as surrogates for a long-term outcome.

Large AI models increasingly rely on distributed GPU execution, yet it remains unclear whether established failure analysis methods transfer effectively to this setting. We introduce GF-Bench, a PyTorch-based framework and dataset for generating diverse failures in distributed GPU workloads. The framework supports extensible workload and fault-injection logic; based on this design, the current release targets failures visible in multi-GPU communication logs and provides 220 labeled two-GPU executions covering normal runs, CUDA out-of-memory fail-stop failures, and persistent straggler fail-slow failures. Using GF-Bench, we evaluate three sequence-based methods (DeepLog, LogAnomaly, and LogBERT) and the LLM-based LogPrompt on parsed logs, and additionally evaluate LogPrompt on raw logs. On parsed inputs, no method reliably detects both failure classes, and the highest overall F1 score is 0.667. Raw inputs raise the overall F1 score of in-context LogPrompt to 0.983, at 9.0x its average input-token usage and 8.7x its per-run cost. These results expose a substantial trade-off between effectiveness and efficiency: generic log parsing lowers operational cost but discards the distributed execution context required for accurate failure detection. GF-Bench provides a reproducible basis for studying cost-efficient analysis methods and can be extended to additional GPU workloads and failure types.

We propose a fine-tuning method for flow-matching diffusion models aimed at realistic artificial light modeling without the need for a large training dataset. We address the task of controllable interior image editing, where the goal is to turn artificial light sources on or off while preserving the scene geometry, object placement, materials, and visual identity of the original image. To achieve this, we decompose the task into two independent formulations. We introduce the LoRA Direction Training Method, which extracts the pure direction of the LoRA adapter effect in the diffusion model flow field, and we also introduce specialized loss functions to ensure the realism of the inverse transformation. Additionally, the resulting increment map is used for more precise adjustment of the lighting color and temperature.

Multi-agent LLM systems pair a sender with broad context and an executor with a limited local view. We study when a short message improves the executor's next decision, when raw context is preferable, and when a stronger sender helps. Our framework, receiver-relative bounded coordination, expresses message utility as receiver gain minus protocol tax. Compression beats raw context when tax savings exceed losses from omitted information and decoder mismatch. Even Bayes-sufficient compression can fail when a bounded executor cannot use its surface form. A three-stage decomposition separates externalization, absorption, and action closure, explaining how errors remain after the correct content reaches the receiver. Under a single-crossing condition, sender upgrades help above a receiver-burden threshold. Across six benchmarks, the same Qwen protocol raises ContextBench joint accuracy from $0.633$ to $0.775$ but lowers ToolSandbox from $0.889$ to $0.653$. Fixed-message replay reveals closure failures despite correct artifact recovery. These results guide an inference-time selector that improves the accuracy-cost frontier on the evaluated communication regimes.

Mixture-of-Experts (MoE) architectures allow frontier language models to scale to trillions of parameters, but their deployment is constrained by massive memory footprints and memory-bandwidth limitations. Although modern accelerators provide Sparse Tensor Cores (SpTCs) that reduce weight storage and increase throughput through low-precision semi-structured sparsity, exploiting them for MoEs remains challenging because of substantial model-quality degradation and the lack of grouped sparse GEMM primitives. We present an end-to-end hardware-software co-design framework that compresses expert weights into hardware-native, low-precision sparse representations and accelerates their execution on SpTCs. Algorithmically, our framework relaxes discrete semi-structured support selection through continuous reparameterization, enabling differentiable joint optimization with quantized weights under a router-weighted reconstruction objective and scalable expert-parallel compression. Systemically, we develop a custom grouped sparse GEMM kernel tailored to low-precision sparse MoE inference on SpTCs. Across MoE models ranging from 30 billion to one trillion parameters, our framework improves state-of-the-art joint sparse-quantization accuracy by up to 4.35 percentage points while preserving 96.09% of the original model's performance. On NVIDIA B200 GPUs, our kernel outperforms the vendor baseline by up to $1.65\times$, increasing serving throughput by $1.18\times$ and reducing end-to-end latency by up to $4.03\times$. These results establish hardware-software co-design as a practical path toward scalable and efficient MoE deployment.

Backdoor attacks can be implanted in Large Language Models (LLMs) during training, causing unwanted behaviour when a trigger appears in the input. Existing backdoor defences for LLMs attempt to remove the backdoor but inadvertently shift the model's output distribution to benign prompts, which can result in degraded model performance and safety. We propose NEEDLE, a training-free method for targeted backdoor removal. Once a trigger has been identified, our method estimates a backdoor direction and a refusal subspace through activation vectors, then applies sequential weight orthogonalisation to suppress the backdoor while preventing changes in refusal-related representations. NEEDLE requires neither a clean reference model nor the original poisoned training data. Evaluation is conducted across multiple model families and attack types. NEEDLE achieves the lowest mean Attack Success Rate (ASR) among the evaluated defences, including 0% on challenging code injection attacks, while resulting in the lowest KL divergence and minimal changes in capability and safety.

Software that hands branching decisions to a model needs a declared option and a probability it can threshold. Typed decision models, also called System One models, return such probabilities without generating text, while supervised classifiers and generative language models are the established alternatives. One harness sends eight decision-model checkpoints from six families, including the hosted model Jev, and four open generative models from three developers the same semantic requests, and scores trained and zero-shot classifiers on the same workflow, intent, emotion and social-science items. With task labels, a fine-tuned DeBERTa-v3-large has the highest observed accuracy on every labeled benchmark but one. Without labels, no decision model is significantly more accurate than Jev on workflows or intents, but Gemma-4-31B matches it on workflows and exceeds it on CLINC-150 at higher cost and latency. Stated probabilities of generative models become unreadable when replies miss the key format, whereas key likelihoods avoid this but can saturate. A guaranteed 5 percent risk leaves Jev 0.528 of the intent decisions, and an in-scope threshold still accepts 0.310 of out-of-scope requests. On typed-decisions, swapping yes and no flips 50.5 answers per hundred for Jev and at least 16.8 for every generative model tested, against at most 6.5 for four fine-tuned decision checkpoints. Exposure to a benchmark's training data explains the largest lead of an open checkpoint, which vanishes on rater-labeled emotions. An intent-trained first stage escalating to Gemma-4-31B reaches that model's accuracy at about Jev's price. The results yield condition-dependent design rules for automated decision gates.

Multimodal on-policy distillation (OPD) aims to improve visual grounding while preserving the strong reasoning capabilities of language models. Recent multi-teacher approaches combine LLM and VLM teachers to provide complementary supervision. However, directly using a VLM's full predictive distribution entangles its visual grounding signal with its own language prior, preventing the grounding information from being transferred independently. Conversely, increasing the strength of visual supervision can improve perception but may overemphasize visual evidence and degrade language reasoning. To address this trade-off, we introduce LEGO-OPD, which selectively composes factors from a Language Expert and a Grounding expert into One teacher distribution for multimodal OPD. Under a generalized Bayesian formulation, the language expert provides a prior over candidate tokens, while the grounding expert contributes a visual likelihood that updates this prior, rather than transferring its complete predictive distribution. This factorized composition allows language reasoning and visual grounding to be controlled independently. We further introduce adaptive calibration to determine how strongly the visual likelihood should update the language prior at each decoding prefix. Specifically, LEGO-OPD uses the grounding expert's image-induced prediction shift as a prefix-dependent reference, preventing both insufficient and excessive visual supervision. Experiments with Qwen3 models show that LEGO-OPD consistently outperforms the evaluated single- and multi-teacher OPD baselines on both multimodal and text-only reasoning tasks. Moreover, it improves the initial student's visual perception while preserving text-only reasoning.

Distilling the reasoning capabilities of large language models (LLMs) into smaller students is a central challenge for efficient deployment. Current approaches face a fundamental tension: optimizing purely for verifiable task rewards (e.g., via GRPO) leads to reward hacking, where students arrive at correct final answers through flawed intermediate logic, while regularizing with soft divergence penalties against a teacher (e.g., KL-based distillation) dilutes task performance and, critically, allows the student to compensate for severe logical violations at one step with high teacher agreement at others. We argue that this averaging is fundamentally misaligned with the nature of reasoning: a chain-of-thought is only as valid as its weakest link. Motivated by this observation, we formulate reasoning distillation as a constrained reinforcement learning problem in which the task reward is maximized subject to a worst-case constraint on the teacher log-likelihood along every prefix of the trajectory. To avoid the prohibitive cost of dual Lagrangian solvers and the test-time teacher dependence of state-augmented methods such as Saute, we derive an unaugmented constrained MDP whose reward transformation preserves the hard-constraint semantics, admits a low-variance policy gradient decomposition into single-step and long-term terms, and provably satisfies the worst-case constraint almost surely in the penalty limit. Through extensive experiments on mathematical reasoning and code generation tasks, we demonstrate that our method significantly expands the accuracy-fidelity Pareto front. By matching the high Final Answer Correctness of pure RL and drastically reducing teacher constraint violations, we ultimately achieve the highest rigorous Reasoning Success Rate across all evaluated settings.

Selective state space models (SSMs), such as Mamba, S4D, and LRU, are bounded by transition matrix commutativity (A_t A_t' = A_t' A_t) and solvable affine transformation groups (Aff_D of derived length <= 2). Consequently, stacked multi-layer diagonal networks face severe optimization degradation on non-solvable simple groups such as A_5 due to the exponential circuit emulation depth required to simulate non-abelian commutators. We propose Non-Commutative State Space Models (NC-SSM), their real-orthogonal counterpart SO(3)-SSM, and arbitrary-dimension Cayley-SSM, lifting state transitions to compact Lie groups SU(2), SO(3), and SO(N). Via closed-form Euler-Rodrigues maps and rational Cayley transforms, NC-SSM achieves exact norm-preserving isometry (||U_t|| = 1). We introduce pure Hopf-fibration Bloch projective readouts (S^3/{+-1} =~ S^2 =~ SO(3)) to eliminate sign ambiguity, true quaternion parallel prefix scans (9.06x speedup at T=2048), and Identity-Gated Lie SSMs to eliminate sparse syntax phase drift. Extensive benchmarks across 14 experimental regimes show: (1) NC-SSM achieves 100% tracking on S_3, D_4, Q_8 and simple group A_5, where a 3-layer deep diagonal baseline collapses to 6.60% (p = 8.81e-4); (2) Cayley-SO(5)-SSM breaks Klein's 1884 ceiling on symmetric group S_5 (50.92% vs diagonal 5.25%, p = 0.0015, delivering 7.8x variance reduction over SO(3)); (3) SO(3)-SSM preserves Riemannian manifolds across 300 steps (< 3.12e-6 drift, > 580,000x advantage), achieving 0.04 deg dead-reckoning error and active tangent denoising; (4) NC-SSM achieves 74.36% on Dyck-2 and 30.26% on deep AST scope tracking (p = 0.0081); and (5) ablation confirms strict isometry is mathematically necessary for lossless long-range associative memory.

Structured tool calls often fail after only a small number of fields violate a schema or an execution contract. Regenerating the complete object enlarges the action surface and makes repeated repair difficult to audit. We introduce ContractRL, a contract-constrained sequential repair protocol that models verifier-guided JSON repair as a bounded decision process. At each step the policy observes the candidate, typed verifier feedback, JSON Pointer, immutable repair history, and remaining budget; a contract-derived action mask filters malformed or prohibited RFC-6902 operations before a deterministic validator performs the transition. We specify a contract-constrained group-relative objective for patch, retry, and abstention decisions while keeping canonical targets and semantic labels outside the online state until trace freeze. Under identical verifier information, ContractRL attains 0.9362 semantic success with 34.4 generated tokens, compared with 0.9076 and 44.9 tokens for Patch-SFT and 0.9148 and 137.2 tokens for full regeneration over 192 cases per seed and five seeds. Policy optimization improves semantic success from 0.9186 for supervised ContractRL to 0.9375. A separate three-seed paired evaluation against Patch-SFT yields a semantic difference of +0.0396 (95% CI $[+0.0137,+0.0662], p=0.0039$). Feedback, action-mask, budget, and schema-shift analyses connect these gains to localized correction, while adversarial and multi-turn evaluations characterize the remaining failure modes.

Tool-using agents can expose citations and execution logs while leaving a critical association unaudited: whether the claim shown to a user is the claim emitted by the committed execution and supported by the cited source. A valid citation and a valid trace can therefore remain individually well formed while being transplanted across claims, actions, runs, or source versions. We introduce a claim-anchored execution contract that jointly binds the emitted claim, its exact source span, the ordered execution prefix that produced it, and the source version and access state observed by that execution. Each receipt contains an emission anchor that deterministically locates the claim inside a committed answer or claim-bearing action, together with source identifiers, offsets, hashes, quotes, and a domain-separated execution commitment. A deterministic integrity verifier reconstructs these bindings before semantic or task labels are joined. We separate this integrity plane from a pluggable support plane, so structural validity is not used as a proxy for entailment. The contract exposes seven independently testable properties: claim-emission binding, source binding, ordered-execution binding, oracle separation, persisted-object replay, execution-rerun consistency, and version/access binding. Across 1,280 cross-object attacks, the joint contract detects 1,275 substitutions (0.9961). Removing a targeted property reduces its attack-detection rate to 0.0156-0.0625. On an independently adjudicated 384-pair split, the conflict-aware support guard reaches F1 0.8865 and false acceptance 0.0729; on unseen failure families, these rates are 0.8679 and 0.0938.