Training Data

We study length inflation in on-policy distillation (OPD), where student responses can become excessively long and even exhaust the generation budget. We identify \emph{termination-token mismatch} between base students and post-trained teachers as an important source of this behavior. Across Qwen3, Llama, and Gemma, the two models can place their stopping probability on different EOS tokens, even when their declared stopping sets are identical. This mismatch can suppress the student's preferred termination action without reliably transferring the teacher-preferred alternative. We show that aligning the decoding stopping set alone is insufficient, while treating functionally equivalent EOS tokens as a shared semantic stopping action substantially mitigates mismatch-induced length inflation across all three model families. To further understand how termination behavior evolves over training, we study OPD across different K2-Horizon training stages. This stage-wise analysis shows that termination preferences can shift substantially during training, while also revealing a distinct length inflation late in the OPD run that persists beyond termination alignment. Together, these results identify termination mismatch as an important, but not exhaustive, source of OPD length dynamics. We release an implementation incorporating the proposed termination-handling corrections.

Tactile signals provide direct contact and force measurements that are essential for understanding physical interactions and enabling dexterous robotic manipulation. However, tactile sensing requires direct measurement at contact interfaces, making large-scale data collection reliant on intrusive, costly, and restrictive instrumentation. We present TouchSight, a monocular egocentric vision framework for dense full-hand contact force prediction that leverages 500 hours of pressure-glove recordings and extensive hand-object interaction (HOI) data. To address the appearance gap between gloved training data and bare-hand real-world scenarios, we construct TwinTouch-20H: 20 hours of paired visual data in which generative video models re-render gloved recordings as bare-hand observations against new backgrounds while preserving the original measured tactile labels. TouchSight predicts dense force from both gloved and generated bare-hand videos, outperforms prior contact prediction methods on OakInk2, qualitatively generalizes to natural bare-hand egocentric videos from unseen datasets, and improves consistently as glove supervision scales. These results demonstrate that dense tactile signals can be recovered from egocentric vision alone, without tactile instrumentation at capture time.

Ubiquitous time series data across diverse domains enables critical applications in areas such as transportation systems and power grids. Recently, training foundation models on massive datasets to achieve accurate zero-shot forecasting has emerged as a major research focus. However, current studies predominantly prioritize architectural innovations while insufficiently addressing data diversity, often relying on simple data sampling strategies that fail to manage complex data distributions effectively, leading to inefficient use of training data and suboptimal performance. To address this, we propose QUALS, a large-scale time series corpus equilibrium framework. QUALS significantly enhances data efficiency, i.e., enabling existing models to achieve superior performance using only a small fraction of the original training data. Specifically, QUALS operates through two core mechanisms. First, a pattern quantization framework systematically decodes heterogeneous patterns from mixed corpora via vector quantization and uniform binning. Second, a learnability synchronization framework calibrates sampling weights for heterogeneous patterns, bridging the optimization gap between simple and complex motifs to maximize overall training efficiency. Extensive benchmarks demonstrate that pre-training on QUALS consistently achieves superior zero-shot performance, even under substantially reduced training budgets.

Subliminal learning allows language models to transmit behavioral traits through training data with no obvious semantic relationship to those traits, undermining content-based data filtering as a safety intervention. Training data attribution offers an alternative: it identifies the training examples responsible for a given model behavior, independent of their semantic content, and so may apply in exactly the cases where semantic inspection fails. We evaluate three gradient-based attribution methods (GradCos, a contrastive GradCos variant, and EK-FAC) across three models, comparing them against divergence tokens, a strong baseline previously shown to localize subliminal learning (albeit one that requires access to counterfactual teacher models). Filtering at the token level, EK-FAC mitigates a significant part of the effect, the other methods provide little benefit, and all mostly fall short of divergence tokens. Filtering entire samples is less effective for every method, though EK-FAC often gives a stronger signal than divergence tokens in this setting. Success is inconsistent across methods and settings: variants that work well for some model-preference combinations fail for others, and we do not identify a consistent explanation for these differences. Our results suggest that gradient-based attribution can identify data responsible for subliminal learning in some settings, but that some approximations are more reliable than others.

Imitation-learned robot policies are frequently overfit to the visual conditions present in their training demonstrations. Consequently, variations in object color or background appearance often induce substantial performance degradation. A common mitigation strategy is to acquire additional demonstrations in each novel visual context; however, this approach is resource-intensive, requiring repeated access to a robot, a controlled environment, and human operation for every appearance condition to be covered. We introduce ReShoot, a framework that synthesizes visual diversity by re-rendering previously recorded demonstrations under altered appearances, thereby shifting the burden from data collection to generation. A vision-language model captions the scene, edits a targeted attribute (e.g., background, object color, or material), and an edge-conditioned video generator re-renders both camera views to match. The instruction is updated accordingly. The action sequence and proprioceptive trajectory are copied verbatim without relabeling, so each generated episode retains the recorded action and proprioceptive labels. On LIBERO, a policy trained on an equal mixture of recorded and re-rendered demonstrations matches the performance of recorded-only training (96.5% vs. 96.9%). Moreover, the mixed training set improves robustness to scene perturbations on LIBERO-Plus (85.5% vs. 82.3%). Across two physical robotic platforms, deploying ReShoot with 43 and 100 pre-collected demonstrations increased the success rate on recolored objects from 0.0% to 42.9% and 47.5%, respectively, while maintaining performance under the original recorded appearance.

Data attribution methods using gradient similarity are widely used to analyze and select training data for large language models, but what gradient similarity actually measures is debated. Some interpret it as identifying task-relevant skills, while other work reports that surface form is the main factor. We resolve this debate for supervised fine-tuning examples by varying task and answer format independently. Specifically, we render benchmarks in different answer formats, such that datasets can share a task without a format or a format without a task. We find that gradient alignment follows the answer format, as benchmark pairs sharing an answer format align strongly (disattenuated cosine near 0.4), while same benchmarks rendered with different answer format classes show no alignment (near 0.0). We demonstrate that this ordering holds from the earliest pretraining checkpoints through post-training, and across model scales and families. We then analyze the released selections of LESS, a gradient-based data selection method for instruction tuning, and find that each target's selections over-represent the target's own answer format. Hence, we demonstrate that gradient-based attribution methods track format similarity more than task semantics, meaning that such methods, as well as the semantic interpretation of the gradient, should be tested on data where answer format and task vary independently for greater robustness and reliability.

Artificial intelligence (AI) is increasingly integrated into healthcare systems worldwide, yet its successful clinical adoption depends critically on workforce readiness, particularly in low- and middle-income countries (LMICs) where infrastructural and training gaps persist. This cross-sectional study evaluated awareness, attitudes, preparedness, and barriers to AI adoption among 761 healthcare professionals across multiple disciplines and practice settings in Nigeria. Data were collected between December 2025 and March 2026 using a structured, validated questionnaire. Overall awareness of AI in healthcare was high (92.6%); however, objective knowledge and self-reported preparedness remained limited, with 40.9% reporting low or very low knowledge and only 63.0% feeling adequately prepared. Willingness to adopt AI was high: 92.5% expressed interest in training, and 78.7% supported inclusion of AI education in undergraduate curricula. Key barriers included lack of training (84.7%), poor infrastructure (71.1%), high cost of AI tools (61.0%), fear of job displacement (60.6%), ethical concerns (52.9%), and data privacy concerns (52.7%). Significant differences in preparedness were observed across geopolitical zones (chi-square (5) = 24.28, p < 0.001), and awareness differed across professional groups (chi-square (6) = 68.38, p < 0.001). Attitudes toward AI differed significantly across professional groups (F = 3.32, p = 0.003), with professionals who felt prepared demonstrating more positive attitudes (mean = 3.74) compared to those who did not (mean = 3.46). These findings reveal a critical disconnect between high awareness and actual readiness, underscoring the need for targeted training, infrastructure investment, and clear implementation frameworks to bridge the gap between AI technological potential and clinical reality in resource-constrained settings.

Safety alignment teaches large language models (LLMs) to recognize harmful requests and reject risky instructions. Yet aligned models can fail when harmful intent is concealed within seemingly benign contexts. Robust safety therefore requires both knowledge of safety boundaries and \textbf{vigilance}: the ability to detect unusual premises, misleading reasoning, and latent risks beneath surface-level semantics. Vigilance requires models to scrutinize a request's underlying intent and assumptions before acting. To cultivate this capability, we introduce \textbf{cunning questions}, which are not necessarily safety-related but contain misleading premises, atypical reasoning, or subtle inconsistencies. We hypothesize that learning to look beyond such reasoning traps can transfer to safety-critical scenarios. Experiments show that Cunning training improves robustness to out-of-distribution jailbreak attacks and strengthens subsequent safety fine-tuning. Furthermore, augmenting an existing state-of-the-art safety alignment pipeline with Cunning establishes a new state of the art across our evaluated settings, reducing mean ASR across nine backbone--benchmark combinations from 17.40% to 15.05%. Trace analysis after matched safety fine-tuning suggests that safety judgments are more likely to govern responses before harmful planning begins. A conditional theoretical analysis further characterizes when invariance learned from cunning data can transfer to safety-related inputs. These findings suggest that cunning data can strengthen model vigilance and complement conventional safety alignment.

Multi-encoder fusion extends Large Audio-Language Models (LALMs) beyond speech-centric recognition, but selecting encoders via intuition or exhaustive search often introduces redundant representations and inflates an already constrained compute budget. We propose CUES (Correlation-gUided Encoder Selection), a lightweight heuristic that estimates complementarity through task- and category-level Pearson correlations between encoders' performance profiles, scoring a candidate set from single-encoder evaluations alone--without fusion training during selection. Evaluated on the XARES-LLM benchmark with a frozen SmolLM2-135M backbone (LoRA-adapted) via five-fold cross-validation, CUES consistently identifies the same configuration per track from held-out development splits alone, without using test data for selection. For the broad TrackA suite, CUES selects a cross-family trio (Whisper-medium, mHuBERT-147, and Dasheng-base), achieving a 4.3% relative gain over Whisper-medium (0.771 vs. 0.739). For TrackB text generation, it re-anchors on a focused, speech-only pair (mHuBERT-147 and WavLM-base-plus) and actively abstains from adding a divergent encoder, outperforming mHuBERT-147 by 6.3% (0.589 vs. 0.554). Rather than a failure to scale, this divergence is consistent with a diversity--interference trade-off that CUES navigates per track from correlation signals alone: across the evaluated pool, added cross-family diversity tends toward an inverted-U on broad audio tasks but toward steady degradation on text generation, which favors a focused, speech-anchored set.

Open-source language models have a reproducibility problem. Despite releasing weights, training data, and recipes, none of them are provably reproducible due to the non-associativity of floating-point arithmetic. Deep learning frameworks often offer a deterministic execution mode, allowing reproducible operations on the same machines. Unfortunately, this determinism does not carry across hardware such that a user can verify that a released checkpoint was actually produced using the declared training recipe. This leaves room for undisclosed data, injected biases, or backdoors that existing techniques such as proof-of-learning or proof-of-training-data cannot rule out. We introduce a new tier of model transparency, fully auditable, in which every operation on every data sample during training is independently reproducible on heterogeneous commodity hardware with bitwise certainty. By imposing a definite order on the sources of training nondeterminism, GPU kernel reductions, data batch ordering across a data-parallel cluster, and inter/intra-node collective communication, we make it possible to replay any individual step of a large, distributed training run on a single piece of commodity hardware and check it against the published trajectory. Because replaying an entire run on one machine is infeasible, we support this with a collective verification scheme in which many independent auditors each certify individual steps, together covering the whole run. We release Open-1B, a model trained under this regime, together with its full pretraining dataset, every intermediate checkpoint, the training codebase, and the audit harness needed to reproduce and verify any step of its training.

Multilayer Artificial Neural Networks trained via backpropagation are the basic blocks of many, more complex, classification algorithms. Their strength lies in the possibility of realizing, with arbitrary precision, any function. This result comes at the cost of the large number of involved parameters to be optimized. In this work, we propose a mechanism for cooperation, i.e., information exchange among several artificial neural networks, with the goal of reducing model complexity while maintaining performance. More precisely, we consider several "small" agents, i.e., containing fewer parameters than a reference "large" one, that during training share their predictions by incorporating this information into the loss function and thus directly influence weight updates. We consider several strategies for implementing cooperation, e.g., the voter model, majority model, and weighted average model based on an agent's confidence in its prediction. We numerically compare the accuracy of those strategies on several standard benchmarks. Our results support the claim that several small agents can outperform a single large model on a given classification task; the shared signals affect each agent's optimization algorithm by modulating both the descent direction and the step size, converging toward a global consensus. The proposed proof-of-concept significantly reduces the number of parameters to be trained while preserving comparable performance, thereby limiting computational resource usage.

Automatic prompt optimization (APO) improves language-model programs by revising prompts from task feedback, yet it typically holds its training data fixed. Repeatedly optimizing against the same instances confines feedback to weaknesses already represented in those data, leaving related failure conditions unexplored. We therefore view each failure as a dual signal: it indicates both how the prompt should be revised and what new training evidence should be synthesized. We introduce FORGE, a failure-guided framework that co-evolves prompts and training data. FORGE abstracts imperfect executions into reusable failure modes and synthesizes new training data through four complementary mutation strategies. Verified instances are fed back into prompt search, allowing updated prompts to expose the next data needs. Across eight heterogeneous benchmarks, FORGE improves the aggregate score over the unoptimized baseline by 16.52 percentage points and outperforms all evaluated APO baselines. The synthesized data also transfer beyond FORGE: in a transfer study, they improve all nine APO comparisons by 2--9 points and all three GRPO comparisons by 4--8 points under matched optimization budgets. These results establish failures as a shared interface between prompt optimization and data synthesis, and show the benefit of jointly adapting what a model is instructed to do and what it learns from.

Large language models trained recursively on their own or other models' outputs undergo model collapse, in which distributional tails and factual accuracy deteriorate while fluency survives. Prior work diagnoses collapse after training; the actionable problem is screening a corpus of unknown provenance before training. We introduce SynthSentry, a corpus-level, model-agnostic contamination signal requiring no access to the generating model, no generation history, and no synthetic labels. The score is a distributional divergence over three statistics: lexical diversity collapse, n-gram tail truncation, and perplexity variance across reference models. We evaluate on corpora contaminated by small open-weight generators and an instruction-tuned open-weight model under a leave-one-generator-out protocol. A domain-stratified study measures false positives on naturally repetitive human text (legal, clinical, source code). The score ranks corpora by severity with little loss when whole generator families are held out. Per-domain calibration holds near its nominal false-positive budget once covariance shrinkage and a bootstrap threshold replace a naive quantile, which runs four times over budget. A downstream fine-tuning check showed no contamination-driven accuracy deficit at our scale, so whether pruning recovers one remains open; the same run shows over-pruning risk once pruning exceeds the true contamination fraction. We frame screening as a data-curation defense rather than a post-hoc diagnosis and release the scoring toolkit. All results are small-scale; scope is English-language, batch-mode corpus screening. Contamination sources are single-generation or hand-authored rather than recursively generated, so results speak to synthetic contamination generally and not to recursion depth.

Manchu, now critically endangered, was one of the principal languages of the Qing empire (1636-1912), and its extensive archival record is increasingly digitized but remains difficult to search and analyze at scale. Previous work showed that vision-language models (VLMs) trained only on synthetic Manchu word images can reach 87.4% word accuracy on real Qing manuscripts and prints, leaving a substantial synthetic-to-real gap. This study examines how synthetic and real historical training data should be combined for low-resource OCR. Using 60,000 synthetic and 20,306 real historical word images, we evaluate three pretrained VLMs and a compact convolutional recurrent neural network (CRNN) under four regimes: synthetic-only, real-only, joint synthetic-real, and sequential synthetic-to-real training, following a common checkpoint-selection and archival evaluation protocol. Introducing real training images raises the leading configurations to between 95.09% and 96.28% word accuracy, while no synthetic-only configuration exceeds 87.92%. Synthetic supplementation substantially improves all three VLMs, whereas its marginal effect for the CRNN is sensitive to the training objective. Joint and sequential training yield broadly similar archival accuracy under the tested practical pipelines. A compact CRNN also reaches the leading performance range once real images are available, showing that model scale alone does not determine recognition accuracy. Finally, complementary errors among strong recognizers allow voting to raise accuracy to 98.27% without additional training, while an eighteenth-century Manchu dictionary provides a principled rule for adjudicating disagreements.

Synthetic data has become a promising way to scale model training beyond limited human-generated data but it may also induce strong model collapse (Dohmatob et al., 2024), where any fixed fraction of synthetic data prevents model performance from improving under data scaling, leaving a non-vanishing excess risk floor. In this paper, we study how synthetic data affects the generalization of one-pass SGD in high-dimensional linear regression with model shift. We establish finite-sample risk bounds for mixed and two-stage training, separating standard bias and variance from source-mismatch effects, namely fluctuation and persistent drift under mixing and filtered initialization bias under two-stage. These bounds reveal a sharp contrast: mixed training induces strong model collapse, while two-stage training avoids the floor by using synthetic data only in the first stage, showing that collapse is not inevitable under a simple data curriculum. Under a random sketch model, we further obtain scaling laws for both protocols, with tight results for mixed training in the optimization-saturated regime. These laws show that larger models may amplify synthetic-induced degradation under mixing, and quantify how high-quality synthetic pretraining may reduce bias in two-stage training. Finally, we establish an exact finite-sample necessary-and-sufficient condition for two-stage training to strictly outperform real-only training under the same real-data budget and identical real-stage updates. Overall, our results highlight that synthetic data is neither inherently harmful nor beneficial; its effect depends critically on both its quality and the training protocol used to incorporate it.

Data poisoning corrupts training data to degrade a model or to plant attacker-controlled behavior. This study evaluates two representative training-time attacks, label flipping and backdoor poisoning, on MNIST and Fashion-MNIST with three baseline classifiers: Logistic Regression, Linear SVM, and Random Forest. Clean training is compared with poisoning rates of 5%, 10%, and 20% using clean-test accuracy, macro-precision, macro-recall, macro-F1, and, for backdoors, attack success rate. Label flipping caused clear degradation, largest for Logistic Regression and Linear SVM, while Random Forest stayed comparatively stable. Backdoor poisoning reached attack success rates from 0.9667 to 1.0000 on both datasets and all three models while often keeping clean-test performance near baseline. The results separate indiscriminate poisoning, which shows up in standard metrics, from targeted backdoor poisoning, which stays comparatively stealthy while embedding highly effective malicious behavior, and they support security-oriented evaluation beyond conventional clean-test metrics.

When a language model finds a sentence unusually cheap to predict, it is tempting to conclude that the sentence was in its training data. Almost every published test of that inference has had to guess which sentences were in the training data, the members, and which were not. This paper removes the guessing. Two model families, OLMo-2 and Pythia, publish their pretraining corpora, and a public index over those corpora returns the exact number of times any sentence appeared in each. Those counts make three questions answerable directly. The answers form a pincer, closing from two sides. At the duplication levels ordinary text actually has, five models from 1B to 13B parameters carry at most a faint trace of their own exposure. We measure that trace with a design that reads the same sentence through two models, which cancels fluency and quality by construction, and it comes to a rank correlation near -0.08, where -1 would be a perfect relation and 0 none. Where the trace does become strong, above roughly a thousand copies, the two corpora agree on which sentences those are, because they are the famous ones, so exposure can no longer be told apart from fame. Two further measurements show how apparent membership signal gets manufactured. A common way to build a non-member is to change one word of a member. The model does prefer the original, but the gap is the same whether the original appeared once or a hundred times, so what the model is rewarding is the author's word choice, not memory. Above a thousand copies the gap grows with model size on the twelve sentences we can test there, at the same boundary where the pincer closes. And swapping the controls for sentences that differ from the members in register moves a detector from 0.83 to 0.94 AUC, on a scale where 0.5 is a coin flip and 1.0 is perfect separation. We release the sentence banks, counts, and code.

Machine-learning systems are built almost exclusively on a few large Python-orchestrated frameworks, and they inherit those stacks' engineering costs: environments of hundreds of version-coupled packages, separate export toolchains for deployment, and the split between the language research is written in and the language products ship in. We report on the construction and verification of numbat, a machine-learning stack written in one general-purpose language (Zig) with no third-party runtime dependencies. The stack spans tensor computation, automatic differentiation, neural-network modules, mixed precision, multi-GPU training, data loading and monitoring; an SDK exposes it behind a stable, additively versioned C ABI of over 1,400 entry points, with bindings for six languages; and its clinical domain planes encode regulatory requirements as executable acceptance gates rather than documentation. Verifying such a stack is the harder half of building it: a defective training run rarely fails, it converges quietly to a slightly worse model. We treat a widely used reference implementation as an executable specification and verify against it at five levels, from operator gradient checks to an automated trajectory gate against a same-machine reference run - the arrangement our companion study formalizes as a trajectory-level differential oracle. The protocol surfaced ten silent recipe divergences, which we catalog with mechanisms and symptoms. As the acceptance test, we train a 25.9M-parameter detector of the YOLOv8m class from random initialization on COCO 2017 for the full 500-epoch schedule: the exported weights score 0.4956 mAP50-95 under the official protocol, scored by the reference stack's own validator (published endpoint 0.502), with single-GPU step time at parity on identical hardware. Weights, per-epoch metrics and the full run manifest are released.

As LLMs are increasingly used for pre-submission self-review, there is growing demand for feedback that not only identifies weaknesses but also guides authors toward concrete revisions. We study this as Actionable Peer-review Generation and decompose it into two subtasks: diagnostic claim generation and revision suggestion generation. We introduce ActReview, a rebuttal-guided post-training framework that connects paper-specific diagnoses to concrete, grounded revision plans. Our central insight is that author rebuttals reveal plausible actions for addressing reviewer concerns and can therefore provide latent supervision for revision-oriented feedback. From real review-rebuttal threads on OpenReview, we construct ActReview-40K by aligning reviewer weaknesses with author responses and grounding the resulting feedback in localized paper evidence. We post-train Qwen3-8B-Base with multi-task supervised fine-tuning followed by GRPO using candidate-aware, weakness-specific rubric rewards. We also introduce ActReview-Bench, a human-curated benchmark of 1,000 instances for evaluating diagnostic quality and revision usefulness. Experiments show that ActReview outperforms prior specialized review-generation models on actionability and grounding while remaining competitive with strong prompt-based LLMs. Human evaluation confirms improved revision usefulness while revealing a remaining gap in technical accuracy, and additional analyses support generalization to held-out papers and robustness across independent judges.

While many video-game environments (VGEs) have played crucial roles in advancing reinforcement learning (RL), developing novel VGEs or modifying existing ones to support new features, has been a laborious process requiring extensive hand-coding. Here we present PlayTrain, an RL framework that combines the abilities of large language models (LLMs) to robustly generate JavaScript (JS) games from a minimal human prompt, and an efficient pipeline that can run any JS game in a standard 'gym' environment. Not only are recent LLMs particularly good at writing JS code, but the JS format also allows users to easily play generated VGEs, while PlayTrain enables us to train RL agents on the exact same games. We demonstrate multiple use cases of PlayTrain, including cloning well-known Atari and ProcGen games in simple JS, where PlayTrain trains pixel-based agents end-to-end at over 1M agent-decisions per second on a single GPU node; and creating modified versions thereof (e.g., that support novel test sets, procedural generation logics, or game dynamics). Through PlayTrain, we reimagine RL VGE development: all we need is a single JS file, generated and modified through an LLM. We discuss promising future RL research directions that PlayTrain unlocks.

Cultural heritage collections often contain contemporary and historical visual records of the same physical object. Linking these records is difficult because corresponding images may differ in viewpoint, acquisition conditions, color reproduction, framing, resolution, and degradation, while genuine historical images are frequently scarce. This study investigates whether synthetically aged contemporary images can replace or complement missing historical training data in bidirectional instance-level retrieval. Synthetic old-domain images are generated using degradation-oriented transformations. An EfficientNetV2-M model is evaluated on identity-disjoint training, validation, and test sets across three dataset partitions and three training seeds. Mixed real-synthetic training is compared with real-only baselines using proportionally scaled and fixed 300-batch-per-epoch schedules. Complete replacement of genuine historical images reduced bidirectional mean R@1 from 86.56% to 81.27%, showing that synthetic aging does not reproduce the full genuine old-domain variability. Increasing the number of independently generated synthetic variants provided no consistent improvement. Under controlled scarcity, however, synthetic completion improved mean R@1 by 3.69 percentage points at 25% genuine historical coverage and by 2.92 points at 50%, relative to the proportionally scaled real-only baselines. At 75%, the gain decreased to 2.00 points, while performance remained comparable to the complete-real-data reference. Fixed-schedule real-only controls did not reproduce these improvements. The results indicate that genuine and synthetic observations are complementary. Synthetic completion primarily benefits retrieval by extending cross-domain identity coverage rather than by increasing training exposure, with its contribution gradually decreasing as genuine historical coverage increases.

Benchmark scores describe what a checkpoint can do now, but they do not determine how it will respond to the next training episode. We measure this missing state by branching four short, standardized, target-independent micro-interventions from the same checkpoint and recording their effects in a common capability space. Together with current capability, these responses form L-State; its pulse block supports a flexible direct readout and a structure-preserving operator readout. Under smooth local dynamics, the operator construction admits an end-to-end cross-family bound with explicit source- and target-family coordinate heterogeneity. In three-family leave-one-family-out development, both pulse readouts reduce source-standardized MSE by 39.4% relative to capability alone, while separating the best response and direction estimates. On sealed GLM-4-9B, the direct and operator readouts reduce MSE by 71.8% and 78.3%, respectively, and the operator readout raises sign balanced accuracy from 0.366 to 0.754. On sealed Granite-3.1-8B, the direct readout reaches RMSE 0.544 and a development-fitted action-wise selector reaches 0.554, compared with 1.172 for capability alone. A five-family audit finds that the operator coordinate varies by action and family, and that modeling these deviations improves retrospective held-trajectory prediction. Target-independent interventions therefore expose training-response information that current capability misses, with direct and structured readouts covering complementary transfer regimes.

Wallet reputation scores decide who receives an airdrop, who can borrow, and who enters an allowlist across decentralised finance. They almost always begin as hand-written formulas: compositions of clamped logarithmic, linear and square-root transforms over behavioural features, with every threshold and point award set by hand. Such a formula is readable and deterministic, but it is piecewise and non-differentiable, it cannot improve as data accumulates, and it cannot distinguish a feature that is genuinely zero from one its pipeline failed to capture. We present zScore-N, the neural network that replaced ours in production. The formula served as its teacher: calibrated against 5,208,952 wallets sampled across 2019-2024 and verified to reproduce production output to within 2.3e-13, it supplies unlimited labelled training data at zero label noise. The trained network reproduces it to 0.58 points RMSE on the 1000-point scale (R^2 = 0.99997), against 2.25 for gradient-boosted trees and 28.04 for linear regression on identical features and splits. Trained with missing-value masks against uncorrupted targets, it halves the error that incomplete data introduces: at 10% feature-level missingness the formula drifts 51.4 points from its own complete-data output with a systematic -12.5 point bias, while the network drifts 17.9. The network carries the score at production scale, across a population of millions of wallets spanning six orders of magnitude in size and activity.

Normative age models are trained to predict chronological age in a nominally healthy cohort. Applied to patients, they deviate, and the gap between predicted and chronological age is read as disease risk. Here, we attribute the disease-related effect size of the age gap directly to individual training samples, rather than using a prediction-level loss as the attribution target. For Cohen's $d$, the resulting closed-form influence functional, validated against leave-one-out retraining, ranks training samples by their effect on held-out case-control separation. Across four diseases and two biomarker modalities in UK Biobank, removing the 10% most influential training samples raises held-out disease-related effect size in every seed. It more than doubles the metabolomic-age effect for type-2 diabetes and raises the brain-age effect for multiple sclerosis by roughly a third. Random removal leaves effect size flat even at 50% removal, confirming the gain comes from which samples are removed, not how many. Flagged subjects carry subclinical cardiometabolic burden that diagnosis-based exclusion misses, on markers the model never sees. For type-2 diabetes, where the method gains most, the marker recovered is HbA1c, the standard measure of blood sugar control. We release pyinfluence, our influence-function package, for reproducibility and reuse.

On-policy distillation (OPD) combines student-generated rollouts with dense token-level supervision from a teacher. Existing work has mainly studied its algorithmic behavior, leaving the role of training data unclear. We examine this role at the data-minimal limit by training on a single query. One-shot OPD keeps improving for hundreds of steps and recovers most of full-data OPD's gain across task domains and model families. We explain this result through the states visited during training and the rate at which the student aligns with the teacher. We measure \emph{state coverage}, the fraction of the states full-data OPD visits that a query set's rollouts reach. A single query already reaches $71.5\%$, most of it within the first 100 steps. Adding semantically distinct queries raises coverage and validation accuracy together, until 16 queries reach $98.9\%$ and match full-data training. Yet alignment slows at a similar pace whether OPD trains on one query or the whole dataset, and even a fixed set of states takes hundreds of steps to absorb. OPD is therefore data-overfed but algorithm-starved. Its rollouts quickly expose broad supervision, while the student absorbs that supervision increasingly slowly. The state-coverage result extends to multi-teacher OPD, where 16 semantically diverse queries per domain match full-data MOPD. As a further stress test, content-light templates and off-domain WildChat queries also approach the real-query baseline. Task content and induced state coverage can therefore come apart. We hope these findings direct future work toward the step efficiency of OPD, and prompt a re-examination of the data and the mechanisms behind its recent successes in frontier post-training.

We introduce LLaDA-Image, a unified framework that pairs a 6B Diffusion Transformer (DiT) trained from scratch with a frozen vision-language understanding module built on the LLaDA2.0-Mini diffusion language model backbone. Instead of relying heavily on paired image-text data from the beginning, we first build a strong visual generative prior through image-only pre-training and mid-training. The generation pipeline comprises 220M samples, 98 of which are real images. For efficient and scalable optimization, we use parameter-free RMSNorm throughout the DiT together with the Muon optimizer. The resulting unified model produces highly photorealistic images while accurately following fine-grained editing instructions. We further distill LLaDA-Image into LLaDA-Image-Turbo, enabling fast inference in 2-4 sampling steps. On Qwen-Image-Bench, LLaDA-Image achieves overall scores of 53.53 and 53.38 on the English and Chinese tracks, respectively, setting a new state-of-the-art among open-source models on both tracks. To support further research on capable and efficient generative models, we release our model weights, training code, and detailed recipes.

Learning generalist policies for robust bimanual manipulation is bottlenecked by the scarcity of high quality large scale human demonstration data. In this work, we release 1,500 hours of diverse bimanual manipulation demonstrations covering everyday household tasks, and use this comprehensive corpus to train XR-2, a powerful vision-language-action (VLA) model. Enabled by a purpose built high throughput data pipeline and a carefully designed multi stage training paradigm, XR-2 attains strong manipulation performance in our systematic experiments while retaining favorable training efficiency and high data utilization. We further study two critical scaling axes: varying the amount of expert demonstration data, and post training on DAgger correction data from real time human interventions. In both settings, task success rate improves steadily over the data ranges we probe, exhibiting a clear consistent scaling trend at our current data scale. These results validate both the learning capacity of XR-2 and the promising scaling properties of the released dataset, which we open source to support reproducible research on bimanual robot manipulation learning.

Proteins are fundamental to biological processes, with their function determined by the complex interplay between the amino acid sequence and the three-dimensional structure. Developing generative models capable of understanding this intrinsically multi-modal relationship is crucial for fields like drug discovery and protein engineering. Existing models often rely on a multi-stage training process where autoencoders that tokenize data into latent representations are trained in a first stage. Secondly, a generative model is trained on the latent representation of the autoencoder(s), i.e., generative modeling in a latent space. We hypothesize that this multi-stage training is not necessary to obtain performant co-design models and thus present SimpleDesign, an effective multi-modal protein design model trained directly in the data space. SimpleDesign leverages a single-stage end-to-end objective that combines discrete cross-entropy for sequences and a regression objective for structures. In order to effectively model the difference in sequence and structure modalities, we develop a Mixture-of-Transformer architecture that allows modality-specific processing while keeping global self-attention over both modalities. We train SimpleDesign on over 2M sequence-structure pairs achieving strong performance across co-design and unconditional sequence/structure generation benchmarks.

Pretraining data for Armenian, a morphologically rich and low-resource language, is scarce, and no open Armenian LLM has been released with the data and recipe needed to reproduce it. To address this gap, we curate and release two datasets. ArmWeb is an extensively validated corpus of 4.37M Armenian news documents. ArmSTEM is a parallel English-Armenian collection of 373K math and science problems with step-by-step solutions, translated into Armenian and verified through both answer-preserving LLM judgment and human evaluation. Continued pretraining of Gemma-4-E4B on these datasets yields arm-gemma-e4b, which outperforms every existing open Armenian model as well as its unadapted base, and is the first open Armenian LLM with complete training data and recipe. Our ablations show that news-only continued pretraining improves fluency while eroding knowledge, a pattern we also observe in existing Armenian models, and that a small share of verified translated STEM data reverses the loss. We further find that the largest public Armenian corpora overlap web-derived evaluation panels heavily, including a train/test self-overlap inside FineWeb-2. We openly release all data, models, and code.

Vision-language models (VLMs) have achieved remarkable performance by leveraging complementary information from large-scale image-text pairs. However, missing-modality inputs are commonly encountered during real-world deployment, often leading to significant performance degradation. Existing methods primarily enhance model robustness by learning modality compensation strategies from source training data. However, their reliance on source training data makes them difficult to apply when original data are unavailable due to privacy, storage, or accessibility constraints, such as clinical applications and personalized AI services. This raises an important yet underexplored question: can VLMs be efficiently adapted at test time for visual recognition with missing modalities without accessing source training data? To this end, we propose Test-Time Logit Prompting (TLP), a lightweight source-free test-time adaptation framework for visual recognition with missing modalities. To address missing-induced prediction shifts, TLP optimizes logit prompts with uncertainty-aware adjustment and modality-complete consistency regularization, adaptively adjusting prediction confidence while preserving semantic consistency. Extensive experiments across diverse vision-language benchmarks demonstrate that TLP consistently enhances recognition performance under missing-modality scenarios, achieving up to 8% improvements while requiring only hundreds of tunable parameters and a few test-time optimization steps.

Instruction-tuning data are judged by quality metrics, and tuned models are judged by benchmarks, but both judgments pass through an output interface: the surface format in which an answer is written. Using gradient signatures across 12 tasks, four semantically equivalent interfaces, three model families, and controlled corruptions, we show that this interface confounds both measurements. Spectral statistics such as effective rank are provably invariant to interface rotation and empirically blind to semantic corruption, while the direction of the update carries the quality signal. The interface-varying residual is not noise: it identifies each unit's own target task perfectly across all three families. Capability itself is stored relative to the training interface: a skill that raises accuracy by more than 40 points under the training format can be nearly invisible under every other, and correcting a single generation budget flips the measured effect of fine-tuning on GSM8K from a gain into a large loss. Pre-registered interventions delimit where this geometry stops short of control. Data quality and model capability are interface-conditioned quantities, and current practice often reports the interface instead of the content.

A compressed student has two shapes that need not agree: the weight it deploys at inference and the weight family its training can reach. We show that a state-of-the-art weight-inheritance distiller, Low-Rank Clone (LRC), deploys a full-width student MLP but ties training to a teacher-induced slice, leaving 62.5-81.4% of each deployed matrix's independent linear degrees of freedom unreachable-paid for at inference, never trainable. Our principle is one line: train what you deploy. From the identical LRC warm start, we make the training object the entire deployed matrix, with no change in deployed shape, deployed parameter count, or inference FLOPs, via two mergeable realizations (Dense-LRC and CORE-LRC) that both collapse to one deployed weight. This recovers stranded capacity: taking the stronger realization per teacher, +2.36/+2.71/+10.45 Avg9 over matched-budget plain-LRC baselines across three teachers (Llama3.2-3B, Llama3.1-8B, Qwen2.5-3B), with the largest gain on the widest teacher (Qwen), where it reaches the original recipe's approx. 20B-token accuracy at 10B tokens (2x token efficiency); there the strictly same-lineage arm still recovers +6.39, the fully controlled figure. Controls strongly support attributing the gain to the enlarged reachable set, rather than to added parameters or the recipe. From approx. 10B distillation tokens plus a short SFT, a half-parameter 1.5B student matches its approx. 9T-token teacher's 9-task macro-average, within evaluation noise and with a residual MMLU deficit, and a 2.7B student beats Meta's own official compression of Llama3.1-8B at ~900x fewer compression tokens (a token count under unmatched recipes, not a compute claim). All results are from single-seed runs on the LRC backbone.

Patient exams in the cancer diagnosis and staging process typically generate several whole slide images (WSIs). One of the initial steps in training models on WSI data is identifying one or a few slides containing tumor or other diagnostic biomarkers necessary for downstream prediction tasks such as estimating recurrence risk or progression-free survival. This step requires tedious manual curation by experienced pathologists. Many published datasets make the artificial assumption of 1 slide per patient. Alternatively, all slides per patient may be used for model training, which may dilute the signal from the few slides containing tumor or other relevant information. In this paper, we present a pipeline to overcome these challenges using publicly available WSI foundation models (FMs). Our evaluations show that ranking WSIs based on predictions from zero-shot classification using WSI FMs accurately identifies slides with the most tumor, indicating that WSI FMs contain sufficient morphology signal to automatically triage slides. We also present a formulation for ranked evaluation to benchmark FM performance in slide triage. We show, on multiple datasets, that tumor slides are identified in the top-2 ranked slides for patients with up to 43 slides.

Diffusion models and recursive reasoners are both iterative, but they carry information across iterations differently. We add a persistent hidden state to a diffusion denoiser and remove its timestep conditioning, leaving a single shared update that can be run to arbitrary depth. The result is an anytime solver: accuracy keeps improving with inference depth far beyond the rollout lengths and backpropagation window used in training, reaching 99.90% exact solve on Sudoku-Extreme. We also obtain 98.93% solve rate on Maze-Unique. Surprisingly, progressive denoising is unnecessary at inference: holding corruption at its maximum by replacing every non-clue variable with fresh Gaussian noise at each step retains near-perfect solving and converges to stable solutions. This simple noise-injection mechanism enables a single trajectory to efficiently explore the solution space and settle on the correct answer without parallel rollouts, candidate selection, or external verifiers required by prior reasoning models. Nonetheless, ordered annealed corruption remains critical during training, which suggests that diffusion's primary contribution to our anytime solver is not a sampling procedure at inference, but a denoising training curriculum.

Do multilingual language models transfer factual knowledge across languages during continued pretraining, or do they mostly recall facts learned directly from the target-language data? To answer this question more reliably, we propose an intervention-based framework: starting from an English-pretrained model, we continue pretraining on Persian data from which specific facts have been systematically removed at varying levels of granularity. We construct SIFT, a resource of 500 triples across 20 relations, stratified by the cultural origin of each fact's subject into general (globally prominent) and Persian-related entities, designed for both systematic fact removal from training data and evaluation, with natively written Persian cloze templates. Our results show that fact transfer is very limited: under the strictest removal condition, a large majority of English-acquired facts fail to transfer into Persian. We further show that sentence-level co-occurrence removal is insufficient to eliminate fact signal, and that easier (randomly selected) negative candidate sets substantially inflate apparent transfer by rewarding shallow associative heuristics, while performance on a harder candidate set that allows for less reliance on heuristics is much lower. Finally, we show that source-language entity frequency has a large influence, with Persian-related facts, which are orders of magnitude rarer in the English corpus, hardly transferring.

Despite their strong performance, large language models remain highly sensitive to prompt formulation. Prior work addresses this through refined data construction or through dedicated robustness objectives. We reproduce and compare these strategies under controlled conditions, and measure how effective they are in addressing models' prompt sensitivity. We find the current robustness fine-tuning methods improve over standard fine-tuning and in-context learning, but the best-to-worst prompt gap remains as high as 40-57% of performance. Moreover, the recent robustness-enhancing methods we test - CoIN for contrastive alignment and PPCL for consistency regularization - often fail to outperform the simplest data construction strategy: training on one template per batch. Our diagnostics explain these results. The auxiliary objectives move the quantity they penalize, but do not generalize beyond it. Additionally, data construction strategies differ due to the conflicting signs of per-template gradients on 57-64% of parameters. Thus, batches that mix formulations force the optimizer to reconcile competing updates instead of finding a shared, prompt-agnostic one.

Multimodal large language models (MLLMs) are increasingly provided with contextual evidence in heterogeneous forms: as a text passage, as a rendered image of the same passage, or as both together. However, it remains unclear how consistently these surface forms are processed, especially when the evidence conflicts with the model's parametric knowledge. We study modality robustness under knowledge conflict across 13 MLLMs and two datasets, and find them far from robust. (1) Contrary to common belief, models favor a context that contradicts parametric knowledge more readily in image form than in text form; (2) when a contradicting text and image are presented together, the preferred modality is essentially arbitrary, varying with input order, model, and dataset. We further demonstrate that this instability has practical consequences: it degrades performance in multimodal RAG and can be exploited by adversarial attacks. To alleviate this brittleness, we examine several simple techniques---prompting, steering, supervised fine-tuning (SFT), and direct preference optimization; the majority prove ineffective, whereas SFT achieves moderate success. We therefore call for greater awareness of this inconsistency and argue that it is fundamental, demanding attention at multiple training stages.

Generative models have become central across science and industry, from image and text synthesis to the design of molecules and materials. Quantum generative models are considered one of the most promising applications for quantum computers, since a quantum circuit naturally produces samples from the distribution it encodes, and for suitable circuits that distribution is believed to be hard for any classical computer to reproduce. A leading strategy trains these models on a classical computer and reserves the quantum device for generating samples at deployment. This is possible when the training loss can be evaluated on a classical computer. A prime example is the maximum mean discrepancy (MMD$^2$), a moment-matching loss that compares the model and the data through their Pauli-$Z$ correlations. Research so far has asked whether such models can be trained and whether their sampling is hard; whether minimizing such an objective yields a model that generalizes, rather than one that merely reproduces the training statistics, remains poorly understood. We benchmark a broad set of quantum and classical generative models by direct sampling and show that models trained with a moment-matching loss generally show worse generalization than the likelihood-trained models. We show this on two application-inspired datasets: first a cardinality-constrained dataset at up to $30$ qubits and second a dataset of genomic single-nucleotide variants, whose valid set is the observed data. These results indicate that a converged moment-matching loss is not a reliable measure of generalization, and that train-classical, deploy-quantum workflows will need approaches that target generalization directly, leaving open whether better training objectives suffice or whether the model architectures themselves must change.

Industrial post-training is a brownfield regime. Teams inherit a deployed checkpoint and must land targeted improvements under fixed compute and mixture budgets without regressing the rest. The maintained artifact is increasingly dataware: behavior governed by a curated post-training mixture, updated via bounded mixture patches rather than clean-slate retraining. From an industrial code-generation improvement effort, we offer a maintainer's perspective on why this work is hard in practice, distilling three recurring challenges, zero-sum mixture design, yield as the binding metric, and end-to-end integration under uncertainty, and arguing that progress depends less on one-off recipes than on an engineering discipline for programming dataware. In our case study, interventions that raised the conversion of teacher distillation into usable training data increased accepted supervision by 2.84 times while using the same solution teacher and four solution attempts per candidate problem. In our primary evaluation, the yield-engineered patch improved CodeForces pass@1 by +2.59 points (+3.11 pass@3) and held-out LiveCodeBench v6 pass@1 by +6.11 (+8.05 pass@3), all statistically significant across 16 stochastic evaluations of each benchmark from one fixed checkpoint per condition, with internal AIME and MATH regression suites within tolerance.

LLM training is increasingly vulnerable to silent data corruption (SDC), yet existing protection methods largely treat Transformer computations uniformly because their vulnerability remains poorly understood. We present the first systematic characterization of SDC vulnerability across major computation interfaces in both the forward and backward passes of Transformer training. Our analysis reveals two distinct error propagation mechanisms: forward-pass vulnerability is highly location dependent, with faults on the Q/K path producing persistent training deviations, whereas backward-pass vulnerability is largely governed by gradient exponent distributions rather than computation locations. Motivated by these observations, we propose TrainSDC, a characterization-guided protection framework consisting of Q/K-path recomputation, residual-gain monitoring, and exponent-aware gradient scaling. Experiments on Llama 3.2-1B and Qwen3-0.6B show that TrainSDC maintains training behavior close to fault-free execution under both sparse and dense fault injection while introducing only 1.65%-6.76% runtime overhead.

We estimate the conditional population-risk curve of a realized smooth nonconvex gradient flow from the training sample. Flow approximate leave-one-out (Flow-ALO) propagates a deletion response and evaluates omitted observations at approximate deleted paths. The risk-curve error decomposes into response approximation, exact-LOO fluctuation, and deletion-to-full risk transfer. On each fixed finite horizon, bounded centered training-loss gradients, a one-sided Hessian lower bound, locally Lipschitz Hessians, and a strict tube-closure condition yield an explicit $(n-1)^{-2}$ bound for the deletion-response error. Bounded evaluation-loss gradients transfer the deletion-response bound to the score without requiring the Hessian to be invertible. Direct first-order jackknife cancellation and exact-LOO concentration control deletion-to-full risk transfer and fluctuation, respectively, completing recovery of the conditional population-risk curve. For bounded smooth two-layer mean-field networks training both layers, the score-error bound is uniform in width.

Advances in machine learning (ML) have created new opportunities to complement traditional operations research (OR) methods. In particular, transformer models can capture complex interactions in token sequences by mapping tokens into a high-dimensional embedding space and propagating contextual information via attention. This makes them a candidate to model non-permutation flow shop scheduling with secondary resources as a next-token prediction task, where tokens represent job-machine-secondary resource tuples. For training, mixed-integer linear programming (MILP)-generated schedules are tokenized and used as next-token prediction data. During inference, partial token sequences (prefixes) are randomly generated and completed by the trained transformer through constrained decoding. A computational study is conducted on a flow shop with 8 jobs, 4 machines, and 3 secondary resources, where jobs are selected from a fixed pool of 20 jobs that is sampled during training and provides the candidates during prefix completion. The transformer achieves better solution quality (smaller makespans) compared to a genetic algorithm (GA), the NEH heuristic, and random search. It is outperformed only by the MILP model and the iterated greedy (IG) heuristic. The study concludes that transformer models can, to some extent, learn patterns from MILP-optimized non-permutation flow shop schedules and that transformer-based scheduling represents an interesting direction for future research, particularly in settings with a fixed, recurring job set.

Language model pretraining has become almost synonymous with prohibitive cost, placing it out of reach for much of the academic and open-source communities. Although strong open-source efforts already exist, including open-weight models and open-source training recipes, a cost-efficient, hardware-accessible, and open-source pretraining recipe has long been missing. Even at a small scale, training Llama-3.2-3B costs over $1.5M, and reproducing SmolLM3-3B needs over $700K. In this report, we present an open pretraining recipe designed to lower this barrier. Using this recipe, we train a collection of Puro-2B models from scratch on up to 1.4 trillion tokens with FP8 precision on consumer-grade RTX 5090 GPUs. The models in the collection differ in token budgets and selected recipe variants. Our best model is trained at a compute cost of less than $6.9K and approaches Qwen2.5-1.5B performance under our evaluation protocol. This cost efficiency is enabled by a combination of approaches, including hardware selection, low-precision training, hyperball optimization, curriculum model averaging, and the data recipe. Beyond the recipe itself, we provide two additional results. First, across the Puro-2B collection, we derive a Puro Cost Scaling Law that relates training cost to average model performance; the fitted law suggests that about $4.4K, less than $5,090, is sufficient to reach the performance of Qwen2-1.5B. Second, as an end-to-end case study, we examine how pretraining data curricula shape downstream performance after post-training. Such controlled studies are enabled by having access to the full pretraining pipeline rather than model weights alone. We release the full training recipe for Puro-2B, including data, code, and model weights under Apache 2.0 at https://huggingface.co/collections/thu-pacman/puro-2b.

As large language models (LLMs) continue to advance in coding capabilities, their potential in cybersecurity has drawn increasing research attention, with closed-source LLMs (e.g., Mythos) delivering advanced cybersecurity capabilities. However, existing open-source efforts remain limited: frontier open-weight models do not provide reproducible cybersecurity training solutions, open-source training solutions focus on isolated tasks and lack scalable agentic data, and scaling agentic rollouts requires strong domain priors. In this work, we introduce \textbf{CyberFactory}, a unified open-source framework that connects data construction, trajectory synthesis, and model training across proof-of-concept (PoC) generation, vulnerability patching, and cybersecurity question answering (CyberQA). CyberFactory transforms public vulnerability artifacts, including CVEs from the wild, into executable and verifiable task instances. It further uses a reusable vulnerability-analysis skill to guide the teacher through source inspection, problem solving with domain prior, and evidence-based validation. The resulting supervision is agentic: the model interacts with tools and target environments and revises its solutions according to execution feedback. Using these trajectories, we train and release \modelname\footnote{\emph{Aegis} is, in Greek mythology, the protective shield of Zeus and Athena; the name reflects the model's defensive, security-oriented purpose.}, which internalizes the skill-guided procedure without requiring the skill at inference time. On CyberGym, \modelname reaches 52.4% Pass@1 under a one-hour budget, improving over its Qwen~3.5 base model by +22.8 points and outperforming the evaluated general-purpose backbones under the same scaffold.

Measuring training data influence consistently across language model pretraining is challenging. It is difficult to select downstream tasks or validation sets representative of a model's general capabilities, and reliance on task performance at intermediate checkpoints complicates comparisons across training. We propose a measure of training data influence that does not require selecting a downstream task or validation set as the attribution target. Specifically, we define an example's influence by how much its gradient update reduces the squared distance to the final parameters of a given pretraining run, and estimate this quantity from intermediate checkpoints without retraining. Applying the method to 18 configurations from the Pythia and PolyPythia suites, we find systematic temporal changes in influential data. Early in training, literature-related data are more strongly aligned with the trajectory toward the final parameters, whereas STEM data become more strongly aligned in later stages. This qualitative crossover is broadly consistent across model configurations. Our results provide a tractable trajectory-level view of how influential data change throughout pretraining, complementing influence analyses defined with respect to specific downstream tasks or validation sets.

Augmentation can corrupt a training example when an image and its annotations receive different random changes. A crop must use the same coordinates for the image, mask, boxes, keypoints, stereo views, video frames, or volume. Code paths that choose these values separately can silently misalign the data. AlbumentationsX keeps the transform list, probabilities, annotation settings, and random seed in one Compose object. Each call chooses random values once and applies them to every supported part of the training example. The library keeps each object's mask, box, and label together and lets projects add their own transforms. It can also save the pipeline definition, show what happened in one call, and run that call again. The examples place Compose after files have been decoded into arrays and before PyTorch groups examples into a batch. AlbumentationsX executes the declared transforms. Practitioners still decide whether a flip, crop, color change, or other operation preserves the correct label for their task.

Thinking Mode Fusion (TMF) enables large language models to support both concise responses and long-form reasoning by unifying a non-thinking mode and a thinking mode within a single model. However, its training dynamics, including the \emph{data ratio} and \emph{training schedule} between the two modes, remain underexplored. In this work, we present a systematic study of TMF by analyzing the effects of the training schedule and data ratio between thinking and non-thinking modes. Focusing on mathematical problem solving, we construct a benchmark with multiple thinking-to-non-thinking data ratios and three training schedules. Our results reveal an asymmetric interaction between the two modes: increasing the ratio of non-thinking supervision reduces the accuracy of the thinking mode. We further show that different training schedules modulate this trade-off and that the optimal schedule depends on the data ratio. Finally, we quantify a negative correlation between non-thinking and thinking mode supervision, highlighting an inherent tension between these two modes. These findings provide practical guidance for designing effective TMF training settings. All code and data are released to support further research at: \href{https://github.com/caocongfeng/Fusion-Bench.git}{\textbf{Fusion Bench}}.

Establishing the provenance of a language model---including its base checkpoint and possible overlap in training distributions---is a governance challenge that metadata alone cannot resolve. We introduce a training-free fingerprint based on the top-$k$ vocabulary projections of late hidden states elicited by 250 fixed knowledge probes, compared using Jaccard overlap over decoded token strings. We evaluate the method on 32 open-weight models from nine families (0.6B--32B) with documented relationships. (1)~A \emph{similarity ladder} broadly follows model relatedness: independently trained models on identical data score 0.48 raw (0.35 vocabulary-corrected), followed by shared-base fine-tunes (0.39/0.33), same-developer relatives (0.38/0.28), and models with no documented relationship (0.22/0.17). This identical-data signal persists across three organizations, two tokenizer families, and two architecture classes, and emerges within the first 1% of training before measurable task competence, suggesting a contribution from shared training data beyond capability convergence. (2)~As a nearest-neighbor \emph{lineage-retrieval} method, the fingerprint ranks the exact documented base among the top two candidates for all five R1 distillations (mean rank 1.8, MRR 0.60), including a math-specialized base not identifiable from coarse metadata. (3)~A \emph{depth ablation} shows that lineage group discrimination strengthens toward the output distribution, with AUC increasing from 0.72 at quarter depth to 0.90 at the output; using only the top 5 output tokens retains AUC 0.87. (4)~The fingerprint remains stable under quantization, with Jaccard similarity of 0.92 under int8 and 0.82--0.85 under int4, compared with a maximum cross-model similarity of 0.81 in the calibration pool. We release the probes, code, and fingerprints.

A host of recent studies have evaluated the cognitive and developmental alignment of Foundation Models (FMs). These investigations include evaluations of their correspondence to adult performance across a range of cognitive domains, as well as whether aspects of model training track children's cognitive development. However, using FMs as candidate cognitive models poses significant methodological and conceptual challenges. A key question underlies this effort: under what conditions does behavioral alignment justify treating FMs as explanatory models of cognition? In this paper, we articulate a four-stage inferential framework for evaluating FMs as cognitive and developmental models: adapting human experimental tasks to model-compatible formats, specifying linking hypotheses that map model outputs to human measures, evaluating behavioral correspondence, and comparing across candidate models or manipulations. We clarify the role of linking hypotheses in mapping model outputs to human behavioral measures, identify challenges that constrain alignment claims, and propose principles for theory-driven and comparative evaluation. Throughout, we argue that behavioral fit alone is insufficient. Alignment becomes scientifically meaningful only when embedded within explicit theoretical commitments, theory-diagnostic tasks, and systematic contrastive evaluation across candidate models.

Should a completion model spend extra test-time compute by iterating, or spend a similar parameter budget on a wider one-shot predictor? The answer is easily confounded by denoising curricula, corruption augmentation, capacity, and unpaired evaluation. We study this question in LiDAR semantic scene completion by comparing a one-shot predictor, a parameter-matched wider predictor, and a weight-tied multigrid refiner initialized from the same frozen predictor. The protocol separates coherent region removal, independent thinning, range-dependent attenuation, and additive clutter while preserving exact scene-condition pairing. Across five training seeds and 815 SemanticKITTI sequence-08 frames, the full iterative system improves mIoU over the wide control by 0.911 points under contiguous angular removal, with a 95% moving-block bootstrap interval of [0.804, 1.040] that clears a predeclared 0.5-point practical margin. Under independent 75% thinning, iteration adds only 0.300 points [0.166, 0.436], whereas observation-family augmentation adds 5.975 points [5.662, 6.140]. Neither intervention repairs additive clutter. The iterative system also costs 10.74 ms and 0.75 GiB per frame, versus 6.25 ms and 0.23 GiB for the wide control. These results establish a geometry-conditioned empirical boundary rather than a universal advantage: coherent gaps can justify fixed-depth refinement, broadly thinned evidence is addressed more effectively by training coverage, and spurious evidence requires a different robustness mechanism.

Personalizing language models (LMs) to individual user preferences is essential for aligning responses with diverse goals and backgrounds. Existing methods typically train a separate adapter for each user or learn a reward model whose scores depend on the user. Despite explicitly optimizing for each user, these methods must learn from limited observations and therefore suffer from data sparsity and poor generalization to unseen users and domains. In-context learning (ICL) and Context Steering (CoS) can instead provide more effective personalization by conditioning the base LM directly on user context and leveraging its pretrained capabilities without per-user training. Yet neither adapts the influence of that context across decoding steps: ICL leaves it uncontrolled, whereas CoS applies a fixed steering coefficient and requires two LM forward passes per step. We propose Cautious Context Steering (CCS), which adds a lightweight adapter to a frozen backbone LM to decide at each token whether and how strongly user context should affect generation. The adapter learns this behavior from an oracle context-conditioned LM and preserves the base LM when the context is not helpful. A single CCS adapter trained on only one dataset improves generation quality both in-domain and across four out-of-distribution personalization benchmarks, demonstrating robust generalization to new users and domains. CCS also avoids per-user fine-tuning and the additional context-conditioned forward pass required by CoS, substantially reducing inference cost.

Video identity replacement seeks to transfer the identities of one or more subjects while preserving the motion, expressions, and temporal structure of a driving video. Existing methods largely target single-person settings and often require task-specific structural controls, such as masks or pose representations, limiting their flexibility in general multimodal editing systems. Progress on multi-person replacement is further constrained by the scarcity of paired training data. We present Vorch-IR, a unified framework that supports single- and dual-person identity replacement, with optional background replacement, in a single model. Built on LTX2, Vorch-IR jointly conditions on a driving video, indexed reference images, and a textual editing instruction. The reference images need not match the pose, layout, or spatial configuration of the driving video: their roles as subject or background references are specified through the instruction. Dense visual conditions are fused through self-attention, while a vision-language context establishes semantic correspondence through cross-attention. We further develop an automatic data construction pipeline that synthesizes paired supervision for all four editing settings. Experiments using automatic metrics and pairwise human evaluation demonstrate strong identity preservation, motion fidelity, and temporal coherence across diverse scenarios. A temporal overlapping inference strategy additionally extends the short-clip model to minute-long generation without autoregressive continuation.

Transcutaneous spinal cord stimulation (tSCS) modulates spinal sensorimotor circuits primarily through activation of afferent networks. While prior work has emphasized locomotor performance and spinal excitability, how tSCS affects conscious proprioceptive perception and the extent to which such effects parallel changes in locomotor control remain unclear. We investigated the acute and training-related effects of tSCS on ankle proprioception and gait in unimpaired adults (n = 14), with an independent control group (n = 14) completing identical proprioceptive training without stimulation. Proprioception was quantified using a bilateral robotic assessment of dynamic ankle localization ability (Crisscross), gross motor output using maximum dorsiflexion strength, and gait during normal and tandem treadmill walking using spatiotemporal, trunk-sway, and mediolateral center-of-mass (CoM) excursion measures. Acute tSCS increased ankle proprioceptive error (p < 0.001) while dorsiflexion strength was unchanged (p = 0.30). Gait shifted toward a modestly more constrained locomotor pattern, characterized by reduced step width and ML CoM excursion (p < 0.05). With continued training under stimulation, proprioceptive error decreased and, unlike the control group, the tSCS group showed progressive improvement that persisted after stimulation ended. Sagittal-plane gait measures recovered toward or beyond baseline, whereas mediolateral measures remained constrained, revealing a direction-dependent reorganization of locomotor control. Together, these findings show that tSCS influences multiple aspects of the sensorimotor control loop, disrupting conscious proprioception while reshaping locomotor behavior, and that the nervous system can adapt to altered afferent input through training.

Fine-tuning a large language model on new data degrades what it previously learned. We present Omega-S, a drop-in penalty computed from the weight matrix alone: it needs no previous-task data, no Fisher matrix and no stored copy of the old weights. It is three lines in an existing training loop and adds under 4% to the cost of a step. Retention. On Llama-3-8B with LoRA, fine-tuned from code to prose and measured by HumanEval over ten seeds, Omega-S retains more of the original capability than no regularisation on 9 of 10 seeds (0.173 -> 0.238 absolute pass@1; sign test one-sided p=0.011, Wilcoxon p=0.006), as a retention ratio, 62.9% -> 84.1%. It also beats tuned weight decay on 10 of 10 seeds (p=0.002) and tuned EWC on 8 of 10 (p=0.014), every arm re-measured in the same session. Mechanism, measured rather than asserted. Omega-S is topological by construction, its objective built from Tr(A^3), but we measured which of its four factors actually moves and three do not: their elasticity with respect to the weights is at or below 1e-4, against 9e-3 for the degree-variance term. As implemented, the composite reduces to a penalty on the variance of node degrees, which means row magnitude in square modules and directional alignment in non-square ones. We report this because a method whose name promises one thing and whose gradient does another should say so. We also enumerate the open design choices, including a contrast-preserving construction that does what it was designed to do and makes retention worse on all ten seeds. Repeating an identical configuration, same seed and same hardware, gives a standard deviation of 0.104 in retention ratio. We have not found this quantified for low-rank fine-tuning of language models, and it bounds every seed-paired comparison in this literature, ours included. Code, per-seed results and the full record of negative results are available.

Long-horizon coding-agent trajectories are poorly matched to the credit units available to train on: a single action has no stable value, an episode label merges productive exploration with abandoned directions, and a fixed window cuts where the logging mechanics fall. We introduce collection-time semantic self-segmentation, in which a declarative contract has the acting agent expose its own boundaries while the trajectory is generated. Instantiated with falsifiable causal hypotheses, successive adoptions expose variable-length semantic phases, and no milestone vocabulary, gold patch, environment replay, teacher logits, or retrospective segmenter places a boundary. Because the agent names its conjecture, a reviewer can negate it by name, which lets our protocol manufacture wrong-cause-then-correction transitions that recorded work rarely contains; one collection then yields four supervised targets, including audit supervision from exactly the failed regions an episode label discards. We then ask what survives deleting the declaration. Given the cut points but not the hypothesis, a model attributes action blocks to their governing hypothesis at over twice chance, beating equal-length blocks over the same trajectories (paired sign test $p = 0.0002$), surviving a lexical control and collapsing under label permutation. Asked instead to place boundaries, a code-blind annotator matches 24 of 40 where random placement matches 11.5, while a mechanical test-event rule beats chance at neither end of a strict-to-permissive sweep. The segments are therefore coherent and not cheaply reproducible. Downstream, DPO on 2,551 phase-boundary pairs changes no decision on 91 adversarial held-out items, while four of 60 change on matched-construction items, all wrong to right, where two controls change none: with 1,825 pairs from one generator, the variable to vary next is corpus diversity, not the boundary.

Federated learning (FL) enables distributed optimization and learning across decentralized edge devices while preserving data privacy, but its performance is fundamentally constrained by heterogeneous data distributions, limited communication resources, and energy availability. In practical wireless networks, mobile devices (MDs) often exhibit diverse data and learning objectives, naturally forming clusters of users with jointly trainable models. When devices rely on energy harvesting (EH), stochastic energy arrivals further complicate participation and scheduling under communication constraints. In this work, we study over-the-air (OTA) FL with EH MDs under heterogeneous data distributions, and investigate two closely related learning objectives within a unified framework: one aiming for a more representative global model by reducing data bias, and the other learning more personalized cluster-specific models by exploiting this bias. In the global training mode, cluster information guides energy- and diversity-aware scheduling, ensuring that the scheduled active users provide a more representative aggregate update. In the personalization mode, the same cluster structure defines cluster-level learning objectives and OTA recovery targets, enabling the parameter server to train multiple cluster-specific models through simultaneous transmissions over the wireless multiple-access channel. Numerical results demonstrate that the proposed unified framework improves fairness or personalization, depending on the operating mode, while reducing communication overhead.

Microrobots hold significant potential for various applications, where targeted navigation is a basic requirement. Deep reinforcement learning (DRL) has recently emerged as a powerful paradigm for fully autonomous microrobot navigation. Yet, current DRL-based approaches pay limited attention to learning efficiency and effectiveness, requiring hours to days for model training. Consequently, this impedes both rapid practical deployment and parameter optimization. To address these challenges, we present a learning framework that enables effective microrobot navigation policies to be trained within minutes. In the proposed framework, we develop a fully vectorized simulator with more than 10,000 artificial vascular environments, parallelizing dynamics, LiDAR-inspired perception, and feasibility checks across thousands of environments to achieve roughly 190,000 transitions per second. To achieve effectiveness in fast training, we propose a task-shaping-regularization (TSR) reward framework. The TSR framework accelerates convergence, improves final performance, reduces action variation by at least 33.7%, and increases obstacle clearance by at least 2.1% across all evaluated scenarios. Results show that the proposed learning framework reduces training time to under 10 minutes, while supporting zero-shot deployment across distinct microrobot types and navigation scenarios. Collectively, this framework can substantially shorten the design loop and accelerate the deployment of autonomous microrobots.

Logical rule induction seeks interpretable rules that transfer across propositional schemas. This requires respecting symmetries: atom naming, example order, polarity flips, and label swap. Enforcing exact symmetry by construction lets one trained inducer scale beyond its training schemas. Our central contribution is a canonical export that decodes a discrete rule from literal scores. It needs no retraining and is exactly equivariant whenever those scores respect the symmetries. We instantiate it on the Neural Rule Inducer, a disjunctive-normal-form (DNF) foundation model that natively respects only example order. We restore the remaining symmetries through architecture, inference, and training. On synthetic stress tests, accuracy on the support labels stays stable at much larger schemas, and rule fidelity on fresh inputs remains above the unmodified model. On real data, accuracy improves most on larger schemas. The exported rule is exact on synthetic full-group tests and on schema-valid real-data tests. This is a mathematical property of the export rather than of a specific model, and we validate it empirically only on the NRI. Enforcing symmetry by construction turns this small-data pretrained model into a reusable, interpretable inducer that transfers to larger schemas.

We introduce DiffusionGemma, an experimental open-weight language model that uses discrete diffusion to generate text at exceptionally high speed. Rather than decoding one token at a time, DiffusionGemma iteratively refines blocks of 256 tokens in parallel, avoiding the sequential decoding bottleneck of conventional autoregressive (AR) large language models. Instead of training from scratch, we obtain DiffusionGemma by fine-tuning the mixture-of-experts Gemma 4 model with 3.8B activated and 25.2B total parameters. Our compute-efficient two-stage training pipeline uses fewer than 10% of the starting AR model's total training token budget. The first stage uses supervised fine-tuning to teach bidirectional denoising, while the second stage combines reinforcement learning with sampler distillation to jointly improve generation quality and inference efficiency. DiffusionGemma establishes a new Pareto frontier for the trade-off between generation speed and model capability. Averaged across our full evaluation suite, it generates around 20 tokens per forward pass and achieves roughly 1,500 output tokens per second on a single NVIDIA H100 GPU, which is substantially faster than AR models even with state-of-the-art speculative decoding. DiffusionGemma also retains the starting model's support for thinking mode, multimodal inputs, and long contexts. Despite diffusion fine-tuning, it remains capable of AR generation with only minor performance degradation, suggesting a path toward hybrid diffusion-AR decoding.

Training large language models (LLMs) to write register-transfer level (RTL) requires large corpora of paired specifications and code, and such data is scarce enough that most public corpora are now synthesized. Synthesis provides scale but not correctness, and in two widely used RTL datasets only 24.4% and 53.5% of pairs pass generated functional tests. This raises the question of how much of such a corpus to keep and which part of it. Correctness alone is a poor answer. A pair that misbehaves in one corner case still shows valid syntax and interface conventions, and complex sequential designs are both harder to generate and harder to validate, so filtering by correctness leaves a corpus of short and simple modules. Correctness is also hard to obtain, since behavior leaves little trace on the surface in RTL, and validating an entire corpus only sorts pairs into passed and failed. We present RTLCurator, which learns a behavior-aware compatibility prior by contrasting each specification with implementations that fail simulation, and calibrates it to a new corpus using a small number of validated pairs. It then constructs the retained subset by balancing alignment, representation coverage, and RTL structural richness. On CodeV and RTLCoder, keeping 80% of the corpus this way improves on training with the full corpus across all reported metrics while validating only 10% of the pool, whereas ranking by the score alone falls below random selection and filtering the whole pool by simulation does no better.