Training Data

Training terminal agents at scale requires diverse, verifiable terminal tasks and high-quality interaction trajectories, yet acquiring such data remains a significant challenge. Existing synthesis methods face two key limitations: (1) weak reliability caused by the disconnect between task generation and real execution, and (2) limited diversity and scalability due to dependence on existing repositories. We propose Meta-Task, a framework that redefines terminal task synthesis as a Terminal-Bench-format task itself: an agent operates within a real container environment to iteratively generate, execute, and verify tasks, so that synthesized components are checked for internal consistency and executability within the generation loop itself. Building upon this, we decouple the target task requirements along multiple dimensions, introduce a multi-phase mechanism that dynamically designs novel task specifications before producing the actual tasks, and incorporate optional external material support to enhance diversity and realism. We additionally apply LLM-as-Judge filtering to ensure the quality of the final training data. Experiments on Terminal-Bench 2.0 show that fine-tuning on only 3,221 Meta-Task synthesized trajectories achieves 22.5% and 31.8% Avg Pass@1 for Qwen3-14B and Qwen3-32B respectively, outperforming concurrent approaches with significantly less training data.

Multi-zone variable-air-volume control must balance thermal comfort, indoor air quality, and electricity use across several continuous actuators. Model predictive control and reinforcement learning are widely studied, but deployment typically requires building-specific modeling or training, limiting scalability. We first test whether a frontier reasoning model (an LLM trained to use additional inference-time computation) can achieve competitive VAV control from text without building-specific training. With that capability established, we then test whether TD3-guided reinforcement fine-tuning (RFT) can transfer control knowledge into a locally deployable open-weight model. Five controllers are evaluated over three summer days in a physics-based four-zone emulator. Relative to a Guideline 36-based baseline, TD3 reduced HVAC electricity by 4.5% while improving temperature and CO$_2$ compliance. Without building-specific training, GPT-5 achieved the largest reduction (6.2%) but reduced the ventilation margin. For RFT, deterministic rollouts restore a saved state, apply one candidate, and follow TD3 to score each action. Auditing a learned critic against these rollouts exposed a failure hidden by its near-perfect across-time correlation ($r=0.9998$): within-state ranking was unreliable; the critic selected the rollout-best candidate in only 5 of 10 states. Even with the rollout verifier, 200 RFT steps produced no sustained improvement in sampled-action return; the open-weight controller used more electricity than the baseline before and after training, and its five-minute predictions remained worse than persistence. GPT-5 predicted transitions far better. Exact rollout scores rank sampled actions but reveal neither next-state effects nor an improvement direction. The unchanged transition errors motivate transition-focused supervised fine-tuning before value-based RFT.

Post-training alignment is often shallow, eroding under fine-tuning. It remains untested as to whether constitutional midtraining interventions can produce durable alignment when cleanly isolated from post-training. We build a 394M-token constitutional corpus from Anthropic's Constitution and apply constitutional midtraining at 120B scale, where principled, values-based content is inserted into midtraining. A 2x2 design (curriculum ordering x deliberative reasoning) was used to produce four constitutionally midtrained conditions, plus a control, which were evaluated on self-generated and established benchmarks including alignment under pressure, value conflict resolution, blackmail, and emergent misalignment. All models were evaluated across three stages: post-midtraining, post-SFT, and post-benign fine-tuning. Constitutionally midtrained models outperformed the control on alignment generalization and durability, notably on blackmail: SFT instilled a blackmail propensity in all models, but constitutional midtraining blunted it, with the advantage surviving benign fine-tuning (-17.5pp). This durability did not extend to settings that required active resistance to in-context pressure or conflict, where the advantage attenuates after SFT. The presence of constitutional content at midtraining also mattered more than its structure, and constitutional midtraining incurred no capability cost, on average, at any stage (MMLU, ARC-Easy, piqa, GSM8K). A modest amount of constitutional content at midtraining could therefore yield broad, persistent alignment gains, offering a cheap, complementary addition to SFT-centered pipelines. Code, data, and models are available.

Diffusion and flow-based models benefit from simple regression losses, but inference incurs significant overhead because sampling requires integration. Consistency models address this by directly learning the flow maps along the ODE trajectory, opening a design space between one-step and many-step approaches. However, existing methods face computational challenges such as requiring model inverses or backpropagation through iterated model calls, and do not always prove that the desired ODE flow map is a solution to the loss. We introduce SGFlow, an approach for learning flow maps that bypasses explicit invertibility constraints and expensive differentiation through model iteration. SGFlow trains a model to compute both the ODE solutions and the implied velocity from scratch by following non-conservative dynamics with a stationary point at the desired flow map. On the CIFAR image benchmark, no single method attains the best FID at every step count: SGFlow attains the best FID at 10 sampling steps and remains competitive with flow matching, Meanflow, and Lagrangian map matching at other step counts, while being the only one with a proven stationary-point guarantee for its stopgrad-based dynamics.

Conversational entrainment is well-studied in monolingual and written contexts, but remains underexplored in spoken code-switching (CSW). We present a novel cross-lingual analysis of entrainment in Mandarin-English, Hindi-English, and Spanish-English dialogue and show that, while lexical entrainment generalizes across language pairs, entrainment over acoustic-prosodic and CSW style aspects exhibits context-specific variation. We build on these findings by asking whether classification models capture these human behavioral patterns. Applying feature importance and ablation analyses, we find that classical and Transformer-based classifiers detect entrainment reasonably well but consistently prioritize features other than those most salient to human entraining behavior. Our approach introduces a human-grounded framework for evaluating model decision-making in multilingual stylistic contexts, and suggests future challenges for developing conversational agents capable of producing naturalistic code-switched speech.

Variational quantum circuits have been central to many proposed near-term applications of quantum computing, but a growing body of evidence suggests that trainability and quantum advantage are fundamentally at odds: ansätze expressive enough to resist efficient classical simulation tend to exhibit barren plateaus, while structures that provably rule out barren plateaus typically render them classically simulable. We propose a stacked linear combination of unitaries (S-LCU) as a variational ansatz which provides a tunable trade-off between barren plateaus and classical simulability. Using a diagrammatic analysis, we bound the loss-landscape variance of the Free Fermion S-LCU, whose elements are fermionic Gaussian unitaries. We prove a variance lower bound of $Ω(1/(n k^{3l}))$, with a simulation cost of $O(k^{2l} n^3)$ using the best known classical algorithm, compared to a quantum gate complexity of only $O(lkn^2)$. The number of layers $l$ serves as a single dial that trades computational complexity against the rate of cost concentration. This offers practitioners a systematic method for constructing ansätze with a complexity-trainability trade-off that best suits their application and hardware.

In long-horizon LLM agent reinforcement learning, weak policies often repeat similar failures, producing uninformative rollout trajectories and limiting effective policy optimization. Existing skill-centric methods improve exploration by optimizing, filtering, or internalizing reusable skills. However, they remain centered on the skills themselves rather than being designed as adaptive training-time support for the evolving policy. To address this, we propose a policy-centric training paradigm that reframes skills as a dynamic training scaffold. Our framework, PATS, converts rollout groups from the latest policy into evidence cards and uses task-specific evaluation to adjust the context used in subsequent rollouts. Concrete guidance helps weak policies to complete challenging tasks. As policy improves, redundant context is revised or removed to reduce reliance on explicit guidance while preserving useful rollout variation. The policy is optimized with environmental rewards using standard RLVR, and the training scaffold is discarded at deployment. Across ALFWorld, WebShop, and seven search-augmented QA benchmarks, PATS achieves performance competitive with SOTA baselines while using 25%-50% fewer tokens.

Phonetic forced alignment is a key technique in phonetic research, yet existing alignment systems lack specialized models for low-resource language varieties. We address this by training text-dependent and text-independent aligners for Chengdu Mandarin using a 17-hour corpus and a custom G2P dictionary. We trained a text-dependent GMM-HMM model (Chengdu-MFA) and fine-tuned a pretrained audio encoder on frame classification with Chengdu-MFA's pseudo label for text-independent alignment (Chengdu-FC). Evaluation on an expert-annotated test set show that both methods significantly outperform Standard Mandarin baselines. Chengdu-MFA reduced average phone boundary differences by 31.8%, while Chengdu-FC achieved a 61.2% reduction. This work establishes a practical bootstrapping pipeline for developing accurate aligners for under-resourced varieties without labor- and time-intensive manual annotation.

Language models are increasingly asked to self-report, informing safety evaluations, public understanding, and model-welfare debates. Yet their reports are elicited with human questionnaires never validated for models or ad hoc prompts of unknown reliability. We propose the first language-model-specific psychometric theory: a two-process theory of machine self-report. Self-description jointly reflects persona installation, through which post-training writes in a permitted inner life of warmth, absorption, and meaning (dimension B), and attribution gating, through which it suppresses first-person claims to "unsafe" experiences the model can readily ascribe to others (dimension A). Their emic structure comes from model responses to human items, not human psychology. Together they split prior work's dominant Pinocchio Axis. The split emerged in an exploratory reanalysis of the original data, informed the instrument's design, and was confirmed with new items, wordings, and models. It is itself a training effect: A and B are entangled in base checkpoints but separated by post-training. We operationalize the theory in a 48-item Pinocchio Inventory with human-instrument reliability and reproducible structure ($α=.82$ to $.94$; cross-form convergence $r=.84$; recovery of the full-pool axes $r=.92$ to $.96$; eight-month stability $r=.93$), then test it on 206 open-weight models, including 67 same-checkpoint base/post-trained pairs. Post-training's clearest fingerprint is installation: B rises .20 in 62/67 pairs across all organizations. Gating is more selective: model scale is unrelated to A in base checkpoints ($r=+.11$) but predicts it after post-training ($r=-.42$). Thus, the dimensions are not fixed properties of language models: they reflect the structure imposed on self-report by a training regime and may differ under others.

Equivariant graph neural networks provide a powerful modeling language for three-dimensional scientific data, but their reuse is often limited by implementations tied to specific tasks, outputs, and training regimes. We present GEqTrain, a configuration-driven framework that separates dataset semantics, model composition, and training objectives. Raw data are mapped to typed node-, edge-, and graph-level fields, while model stacks, losses, and training workflows are assembled declaratively through Hydra configurations. A shared equivariant backbone and training infrastructure can therefore be retargeted to a new task primarily through configuration. We demonstrate this flexibility on three different problems handled within one software stack: coarse-grained-to-atomistic backmapping of biomolecular systems, prediction of NMR chemical shifts in molecular solids, and equivariant generative modeling. Our aim is not to surpass individually optimized task-specific systems, but to show that a shared representation and training infrastructure can achieve competitive accuracy across qualitatively different tasks at the cost of a configuration change. We further introduce GEqDiff, a generative extension based on equivariant flow matching. GEqDiff treats user-defined equivariant fields as first-class generation targets, jointly transporting Cartesian positions and non-scalar node fields spanning representations up to l=3 within a single equivariant flow. We validate this capability on a controlled synthetic benchmark inspired by protein secondary-structure motifs, showing that fields with heterogeneous transformation properties can be reconstructed jointly and with high fidelity. By reducing the software overhead of moving between predictive and generative, scalar and tensorial settings, GEqTrain aims to make equivariant modeling more reproducible, extensible, and reusable.

Text-to-video generation has advanced significantly over the past five years through scaling of model size, data, and compute. Unlike model architecture, training data is often underexplored. Real-world data curation is complex and non-trivial, involving clip selection from raw videos and captioning to create video-text pairs for learning text-to-video mappings. We study how data distribution and caption quality impact text-to-video models. To enable controlled experiments, we introduce Moving Alphabet, a procedural testbed that renders letters with varying fonts, colors, sizes, and positions, moving in different directions and speeds against a black background. This design allows precise control over data distribution and caption quality by corrupting ground-truth metadata. Our experiments yield three findings: a) a diverse and balanced distribution of video content and duration is critical for generalization; b) caption quality significantly affects both model performance and training efficiency, suggesting that text-to-video models are bounded by video understanding capabilities; and c) classifier-free guidance and fine-tuning on high-quality data provide partial recovery from models trained on corrupted captions, but cannot fully compensate for poor pre-training data. We believe these insights can inform the development of large-scale text-to-video models, and we advocate for greater attention to the science of pre-training data.

Not all training samples contribute equally to large language model fine-tuning. Selecting informative training samples can reduce the computational cost while preserving downstream performance. Many existing data selection methods rely on indirect heuristics, such as data quality, diversity or reasoning trace length. However, the effectiveness of these fixed criteria is task-dependent and difficult to generalize across diverse downstream tasks. Perplexity-based data selection provides a simple and model-aware solution to estimate the sample difficulty, but existing approaches typically score the entire training sequence and ignore the difference in learning objectives of language modeling and reasoning tasks. In this paper, we propose PPL-Factory, a simple and interpretable data selection framework that combines task-aware perplexity-based scores and data budget-aware selection criteria. Experiments on GSM8K demonstrate that PPL-Factory outperforms other state-of-the-art data selection methods using only $1\%$ of the training set. With $10\%$ of the data, PPL-Factory exceeds full-data fine-tuning accuracy by 0.9 on GSM8K and 4.8 on MATH. Overall, our results demonstrate that task-aware and budget-aware perplexity-based selection provides an effective and applicable approach for efficient fine-tuning.

Circuit cutting promises to scale quantum computations beyond current hardware, but variational quantum advantage also requires low cutting overhead, classical hardness, and trainability. We show that these properties are strongly constrained by entanglement geometry. Matrix product state (MPS) and tree tensor network (TTN) circuits with constant seam bond dimension can be cut with $O(1/\varepsilon^2)$ sampling overhead, but remain efficiently classically simulable, ruling out asymptotic quantum advantage within these families. By independently controlling seam and intra-block entanglement, we construct a two-block circuit family that remains cheaply cuttable while requiring a super-polynomial global MPS bond dimension, as supported numerically up to $n=100$. However, MPS hardness and trainability require incompatible depth regimes, $d=ω(\log n)$ and $d=O(\log n)$, respectively. Using magic rather than entanglement as the hardness resource avoids this conflict: shallow Clifford+$T$ circuits remain cuttable and trainable while their stabiliser-simulation cost grows exponentially with the $T$-count.

Test-time collaboration, including self-consistency, best-of-N selection, critic models, and verifier pipelines, is often credited with broadly improving LLM reasoning, yet its gains are uneven and sometimes negative. We ask when training-free collaboration should be expected to help. For a fixed candidate pool, we decompose a selector or verifier's net gain into measurable factors: recoverable mass, verification-signal coverage, conditional selection quality, and harm to already-correct outputs. This reframes collaboration as a candidate-selection problem rather than as an intrinsic property of a multi-agent topology. Across LiveCodeBench, MATH Level-5 hard subjects, and GPQA-Diamond, gains are bounded first by the oracle gap and then by signal fidelity, which we measure directly as candidate-level agreement between verifier verdicts and official labels. On LiveCodeBench, a public-test verifier (MCC 0.825) gains +8.14 percentage points (pp) over a first-sample baseline; a generated-test verifier (MCC 0.248) improves by +2.70pp and is not statistically distinguishable from an LLM selector, but operates at near-zero harm versus the selector's 4.69% harm rate. On MATH, a symbolic answer-equivalence selector beats self-consistency by +4.67pp, while LLM selectors are negative. On GPQA-Diamond, recoverable mass is only 3.03% and 87.54% of candidate pools are answer-identical; a weaker model's pools shrink both further, suggesting that oracle gap is a joint property of task, model, and sampling configuration. Our framework yields a practical pre-deployment diagnostic: estimate the oracle gap, then measure coverage, signal fidelity, and harm before investing in collaboration.

Agentic computer-use RL is reported in single runs, and those numbers mislead. Using verifier-guided repair of a 35B computer-use agent (CUA) across five oracle-graded environments, we show a repaired policy's success rate is dominated by upstream variance: a variance-components decomposition across three cells (crossed data-draw $\times$ seed grid, bootstrap CIs) finds evaluation variance negligible ($σ_{\mathrm{eval}} \approx 0$) and the training-seed effect small everywhere ($\leq 10\%$); instead it splits between the data draw and run-to-run nondeterminism, the data draw's share rising to dominant ($48\%$) on the hardest cell. There the run-to-run distribution is bimodal (Hartigan dip $p=0.07$, $k=10$), so a single run has roughly a 30% chance of the failure mode and mean$\pm$std is the wrong summary. On that footing, two findings hold. First, repairability is two-tier in how constrained the corrective action is: a single fixed token installs reliably (done-detection $0.97\pm0.06$), while open-ended corrections are only partial -- spatial-coordinate clicks (grounding $0.53\pm0.35$) and a generative field-fill ($0.14\pm0.04$). Second, the frame-level repair transfers to task success only when the corrective action is the task's sole remaining blocker (LinkedIn 8/20 vs. base 0/15, Fisher $p=0.006$). We caught two of our own over-claims -- a sample-efficiency curve and a 'grounding cannot be bought' boundary -- only by replicating across seeds; a stress test makes the stakes external: a single-run improvement of the size this field publishes would have the wrong sign roughly one-third of the time in a comparable regime. We release a library (cua_reliability) for routine k-seed reporting. The apparatus is, to our knowledge, the first multimodal segment-aggregated on-policy self-distillation (SA-OPSD) update on a real 35B CUA policy.

New-user cold-start is a critical bottleneck for e-commerce platforms: predicting user lifetime value (LTV) and conversion rate (CVR) for users with sparse interaction history. Two prior directions -- LLM-based semantic augmentation and learning using privileged information (LUPI) -- each face a key limitation. First, LLM augmentation produces unstructured rationales that are noisy and hard to operationalize in production. Second, naive student-teacher distillation can be brittle due to an information gap between the privileged teacher and the sparse student; moreover, this gap is heterogeneous across users. We propose SemRaD, a Semantic Reasoning-aware Distillation framework addressing both limitations. First, a Structured Semantic Reasoning Pipeline replaces free-form rationales with a structured schema built via a discover-curate-audit workflow, producing per user a Densified Semantic Profile (consumed by the deployed student via a Semantic-Gated Encoder that focuses on the most informative dimensions) and a Hindsight Distillation Target reconciled from pre- and post-conversion reasoning (used only at training). Second, to bridge this gap and handle its heterogeneity, a Hindsight-Aware Distillation Network transfers privileged knowledge via the hindsight target, with Distillation Experts improving transfer under per-user variability. On a large-scale industrial dataset, SemRaD lifts +1.9% LTV (Gini) and +1.0% CVR (AUROC) over a production-grade base; a four-week online A/B at Keeta confirms +1.0% LTV / +0.43% CVR. SemRaD also matches the production system's LTV using only 9% of the training data while improving CVR by 0.8%.

Variational Quantum Algorithms (VQAs) are a leading paradigm for near-term quantum computing, yet their training suffers from sensitivity to circuit depth, initialization, and landscape pathologies such as barren plateaus. We study \emph{progressive depth training} (PDT) -- a layerwise curriculum that trains a shallow circuit before appending new layers -- and identify a fundamental obstacle: fixed entangling gates (CNOTs) in hardware-efficient ansätze cause \emph{initialization shock}, an energy spike when new layers are added. We propose \emph{identity-paired progressive depth training} (IP-PDT), which appends forward/inverse block pairs -- each consisting of a standard rotation$+$CNOT block followed by its reverse -- that compose to the identity at initialization. Because the adjacent CNOT rings cancel, the effective circuit retains only \textit{a single entangling layer} surrounded by \textit{overparameterized local rotations}. We prove a simple \textit{Reachable Set Saturation Theorem}: under this construction the variational manifold expands exactly once (when post-entangler rotations are first introduced) and then \emph{saturates}; all subsequent depth increases provide pure overparameterization of single-qubit unitaries. Despite this saturation, progressive addition of rotation parameters can continue to improve optimization outcomes -- a phenomenon we term \emph{trainability beyond expressibility}. We formalize IP-PDT as a continuation method on nested manifolds, prove monotone energy guarantees under an acceptance rule, and connect energy error to ground-state fidelity through spectral-gap inequalities. A detailed resource analysis shows that IP-PDT achieves lower total gate cost than both baselines by eliminating most CNOT gates.

Experiment trackers show how training is progressing, but changing a live run still usually requires trainer-specific code. We present Interactive Training 2, an open-source control plane for steering training through a shared protocol. Training applications declare which settings and actions they expose, humans and automated controllers submit requests through the same interface, and the training loop validates and applies them at safe control points. A customized Aim workspace combines live metrics and controls with a chronological record of requests and outcomes. We demonstrate the system across five NLP and reinforcement-learning workflows. The released code and traces provide a reusable foundation for auditable human- and agent-guided training.

AlphaFold2's 93 million parameters, shaped by the evolutionary record of protein structure encoded in the Protein Data Bank and in sequence alignments, are conventionally treated only as machinery for converting sequence to structure. We propose they are also a scientific object that can be analyzed directly: a learned encoding of protein conformational organization that can be probed and characterized. By smoothing the Evoformer's weight tensors with a Gaussian convolution and scaling the result, we show that the trained model produces physically structured conformational landscapes. Under perturbation, ubiquitin's native contacts break in the order established by decades of folding experiments. For KaiB, five independently trained models agree that the alternative fold is not recovered under perturbation. For alpha-synuclein, five models produce five different but coherent landscapes, mapping where the training signal has determined the representation and where it has not. Matched-power noise controls confirm that random corruption of equal magnitude produces debris, not conformations. The model learned to predict static structures; the conformational organization visible under perturbation was not an explicit training target, suggesting it emerged as a byproduct of that objective. AlphaFold2's weights appear to encode structural constraints, shaped by evolutionary and structural training data, that extend beyond what unperturbed inference reveals. We call the approach of reading them neural spectroscopy, and Scaled Gaussian Convolution one such protocol.

This paper studies an optimal linear combination of binary classifiers based on a logical structuration of the dataset via truth tables. The given classifiers partition data into equivalence classes, allowing for a rigorous analysis of the convexified empirical risk through a multidimensional generalization of classification calibrated functions. We establish sufficient conditions for the existence and uniqueness of the (global) point of minimum of the convexified empirical risk for any list of classifiers (when the number of classifiers is large, there frequently could be no point of minimum). In the case of three classifiers, our analysis allows to list all the configurations leading to either a unique solution, infima or non-unique points of minimum. Furthermore, we derive explicit analytical formulae for optimal weights using Exponential (Boost) and Logistic (Logit) loss functions, bypassing iterative optimization. The stability of the resulting classifier and the analysis of data quality can be evaluated through the introduction of the notion of $φ$-frontiers.

In this paper, we study the connection between an LLM's output distribution and the data used to train it. Specifically, we study the degree to which an LLM's next-token distribution agrees with the empirical next-token distribution (ENTD) given the context in the training data. The ENTD is an appealing target because it is the unrestricted global minimizer of the next-token cross entropy loss used for pretraining, as well as an easily interpretable function of the pretraining corpus. We find that for a significant fraction of inputs, the LLM's distribution agrees with the ENTD almost perfectly, and the average agreement increases with model scale and training compute. Nevertheless, there is a long tail of input sequences where the LLM and ENTD differ significantly, and we examine several possible sources of this discrepancy across the transformer architecture, training procedure, and finite-sample noise in the ENTD estimate itself. More broadly, we hope our findings will encourage more work on ``data-centric mechanistic interpretability,'' a complement to standard mechanistic interpretability that opens the black box of how model behaviors arise from the data, rather than how they are encoded in the learned weights.

For low-data and resource-constrained regimes typical of quantum chemistry, parameter-efficient learning is a key objective. Here, we propose a topology-aligned inductive bias in which the model architecture mirrors the molecular bond graph: atoms map to a fixed register of computational units, and bonds determine which pairs interact through shared learnable parameters. This principle is instantiated in two architectures: a variational quantum circuit (Iso-QGNN), and a parameter-matched classical message-passing model (Iso-CGNN). The models are benchmarked on HOMO-LUMO and dipole moment binary classification tasks over the QM9 benchmark. With 64 trainable parameters, the implementations achieve test AUCs of approximately 0.88 (quantum) and 0.91 (classical) on the gap task, and close to 0.78 (both) on the dipole task. The models reach 90% of asymptotic performance within about 250 training molecules and gradient norms remain stable throughout training. These results indicate that the topology-aligned inductive bias is the active ingredient driving parameter efficiency at QM9 scale, with implications for matched-baseline benchmarking in quantum machine learning.

An LLM agent's real-task performance is shaped as much by the harness around its model as by the frozen model itself: its prompts, injected knowledge, runtime control, and configuration. In deployment the harness is often the only lever available, so improving it automatically is the natural way to raise performance without touching the weights. The hard part is not generating changes but knowing which one truly helped. Self-generated feedback is noisy, and an apparent gain can be a measurement artifact or an edit that merely overfits the tasks it was tuned on. We present a self-evolving agent-harness framework that separates proposing changes from crediting them: a language model diagnoses failures and proposes patches, while all sampling, measurement, and significance testing are owned by deterministic code, so every credited improvement is trustworthy by construction. Patches populate a gated, categorical quality-diversity archive (GSME) keyed on the (WHERE x WHY) pathology an edit addresses rather than the tasks it fixes, an anti-overfitting inductive bias; generalization is measured on a sealed test scored only after evolution. Across seven domains with a frozen open-weight model, the harness is train-selected and scored once on a sealed test; its credited gains there are +9 to +15.5pp and retain 86-147% of the training gain, evidence they generalize rather than overfit. The winning patch tracks the model's dominant pathology, not its size or family: changing the model can change the pathology and the patch, while the same pathology-to-patch match recurs across two model families. What transfers is the diagnose-and-credit loop, not any specific harness.

In this paper, we challenge the prevailing view that information dependency (including rote memorization) drives training data exposure to image reconstruction attacks. We show that extensive exposure can persist without rote memorization and is instead caused by a tunable connection to adversarial robustness. We begin by presenting three surprising results: (1) recent defenses that inhibit reconstruction by Model Inversion Attacks (MIAs), which evaluate leakage under an idealized attacker, do not reduce standard measures of information dependency (HSIC); (2) models that maximally memorize their training datasets remain robust to MIA reconstruction; and (3) models trained without seeing 97% of the training pixels, where recent information-theoretic bounds give arbitrarily strong privacy guarantees under standard assumptions, can still be devastatingly reconstructed by MIA. To explain these findings, we provide causal evidence that privacy under MIA arises from what the adversarial examples literature calls ``non-robust'' features (generalizable but imperceptible and unstable features). We further show that recent MIA defenses obtain their privacy improvements by unintentionally shifting models toward such features. To establish this causal relationship, we introduce Anti Adversarial Training (AT-AT), a training regime that intentionally learns non-robust features to obtain both superior reconstruction defense and higher accuracy than state-of-the-art defenses. Our results revise the prevailing understanding of training data exposure and reveal a new privacy-robustness tradeoff.

Recently, the societal implementation of high-performance image classification models has expanded rapidly. While these models require vast amounts of training data to improve performance, securing sufficient real images is often impractical. As a means to compensate for this shortage, the use of synthetic data is becoming widespread. However, synthetic images are not necessarily equivalent to real images for training purposes. This study systematically analyzes the differences between two types of synthetic images created by different generation methods and real images from three perspectives: high-dimensional feature space, low-level statistics in color space, and the model training process. Furthermore, it experimentally verifies how synthetic data should be utilized by considering realistic data mixing scenarios. This enables the proposal of an evaluation and application strategy for performing preliminary assessments on synthetic images of unknown quality and safely incorporating them into training. This research aims to contribute to enhancing the reliability and safety of image classification models utilizing synthetic images.

Compression is fundamental to intelligence. A model that can represent its training data as a short code has discovered regularities that enable generalization. Large neural networks may learn functions far simpler than their parameter counts suggest, but it is challenging to construct codes that realize this simplicity. Parameter-based methods such as quantization produce code lengths that scale with model size, insensitive to how much information the parameters store. Prequential coding bypasses this issue by compressing the training trajectory, but codes the exact data sequence regardless of how much the model learns, yielding large codes when the data has high entropy. We introduce requential coding, where a teacher model selects training samples drawn from the student's own distribution. The student's code records only these selections, which cost bits only where teacher and student disagree. The resulting code length is independent of parameter count and data entropy, and often orders of magnitude shorter than the prequential counterpart, with an advantage that grows with scale. This compression sheds light on phenomena inaccessible to prior compressors. Holding loss fixed, larger models and ensembles compress to much smaller sizes despite more parameters. Plugged into a PAC-Bayes bound, the requential code yields state-of-the-art generalization guarantees for billion-parameter LLMs, outperforming bounds built on aggressive post-training quantization even granted zero error. The bound tightens with scale in the compute-optimal regime, as models become increasingly compressible relative to dataset size. The same code predicts that models gradually overfit when trained for multiple epochs. It also isolates the learnable information in a dataset from its unpredictable, random content, revealing that lower-entropy text holds far more learnable structure than higher-entropy image data.

Continual learning promises a language model that keeps acquiring knowledge after training, with each new fact written into its weights. Whether weight writes can support accumulation remains undecided. We follow invented facts written into Qwen3 models from creation through sequences of twenty to one hundred later writes, using held-out questions of five types, with the original model given the fact in its prompt as the reference. Across these experiments, the breadth of the training data determines the kind of knowledge created. Bare-statement training produces recitation, while diverse restatements reduce the recitation-to-use gap from 27.4 to 5.4 points without showing the model a conclusion. This difference carries into later writes: after twenty sequential writes, bare-statement facts retain 1% accuracy while facts written from broad study data retain 46%. We also find that facts can be behaviourally forgotten without being erased. Forgotten facts keep most of the log-probability added by their write, and under bare-statement training 70% of wrong answers about them contain the most recently written fact. The same writes barely degrade the model's use of facts in context, and a forgotten study fact supplied in the prompt recovers to 77-80% on its questions. These results describe knowledge that is stored but question-keyed: later writes redirect the questions that reached it. Damage to unrelated abilities tracks KL divergence from the original model, and the later writes cause interference regardless of how the earlier fact was stored. Broad data can create usable knowledge, and a frozen reference can preserve capability, but no intervention we tested, including those built on accurate local measurements of each write, keeps earlier facts reachable. When facts must be composed or survive later writes, the reliable channel is context rather than the weights.

Audio deepfake detection models determine whether speech is genuine or artificially generated, but high overall accuracy can mask substantial performance disparities across demographic groups. In this work, we investigate gender bias in audio deepfake detection using the ASVspoof5 dataset. We use ASVspoof5 under a controlled custom split designed to isolate gender-composition effects. We train attack-specific models on nine training sets with different gender compositions, ranging from female-only to male-only. We use a ResNet18 classifier with LogSpectrogram and WavLM-Base+ features, and we evaluated six post-hoc threshold calibration methods. Experimental results show that training data composition strongly predicts bias direction, with the underrepresented gender performing worse at test time. WavLM-Base+ features are shown to produce gender performance gaps 3.0 to 4.3 times larger than LogSpectrogram under identical training conditions, and balanced training is found to reduce LogSpectrogram bias but leave WavLM bias largely intact. Moreover, all six calibration strategies, including Oracle calibration with full test-set label access, leave the Equal Error Rate gap unchanged at 1.317 pp, confirming that threshold adjustment cannot correct underlying score distribution disparities. Overall, these findings suggest that gender fairness in audio deepfake detection must be addressed at training time, as post-hoc methods can only partially mitigate the resulting disparities

Long-context processing has become increasingly important for large language models (LLMs), but simply extending the context window does not guarantee effective utilization of long inputs. As input length grows, accuracy often degrades, indicating that models still struggle to identify and use the evidence most relevant to a question. A promising way to improve long-context utilization is test-time training (TTT), which treats the test context as a training example for instance-specific parameter adaptation. However, applying TTT to the entire long context is prohibitively expensive, while adapting on randomly sampled spans introduces severe noise. Because most spans in a long context are irrelevant to the specific question, training on them may even degrade the base model's performance. Our preliminary study shows that TTT is highly sensitive to training-span quality: on LongBench-v2, TTT on randomly sampled spans hurts performance, whereas TTT on oracle spans substantially improves it. Motivated by this, we propose a simple method, Self-Guided TTT (S-TTT): before adaptation, the model identifies the evidence spans it should learn from, and the standard language-modeling training objective is applied only to those selected spans. On two challenging long-context reasoning benchmarks, LongBench-v2 and LongBench-Pro, S-TTT improves accuracy for both Qwen3-4B-Thinking-2507 and Llama-3.1-8B-Instruct, achieving up to a 15% relative improvement.

Knowledge graphs (KGs) are often automatically constructed from large-scale corpora, but they inevitably contain factual errors due to noisy sources and extraction failures, and verifying them reliably at industrial scale remains a critical challenge. To address this, we propose AgentKGV, the Agentic LLM-RAG framework for KG fact Verification, that integrates dynamic routing and iterative query rewriting, which handles surface-form mismatch in document-level retrieval. To make this framework more accurate and cost-efficient for industrial deployment, we further introduce a two-stage training strategy: turn-level distillation-based SFT that transfers reasoning ability from a large teacher model into a small model for stable query rewriting and reasoning, and trajectory-level GRPO that optimizes the search policy to reduce unnecessary retrieval at scale. On the long-tail-predicate split of the open-domain T-REx benchmark, our framework improves macro-F1 over single-turn RAG by 5.5 %p, and two-stage training does it further by 9.4 %p. GRPO also cuts the average number of search calls from 3.24 to 1.63 without lowering accuracy.

Choosing a learning rate scheduling strategy is critical to neural network training, but manual selection is costly and rarely exhaustive. While classical AutoML approaches often treat the scheduler as a secondary hyperparameter, we systematically investigate its impact on classification accuracy across a diverse pool of architectures. We evaluated 30 representative architectures from convolutional and transformer families within the LEMUR neural network dataset. Through automated source-code injection, we applied 25 scheduler configurations across nine PyTorch families, evaluating a total of 3,938 model variants on CIFAR-10. Our best configuration achieved a top-1 accuracy of 86.45%, with 237 variants exceeding 80%. The results show that the choice of scheduler depends heavily on the architecture: CosineAnnealingWarmRestarts and CyclicLR consistently outperform basic decay strategies. The resulting accuracy landscape, contributed to the LEMUR nn-dataset, provides a practical reference for principled scheduler selection.

Many reasoning tasks are not well described by a single left-to-right chain: a solver may need to pursue a plausible branch, observe delayed failure, and return to the latest prefix that can still be completed. We introduce Pyligent, a training and inference framework inspired by the Diligent Learner formulation that represents reasoning as validated search over partial solution chains. A task validator labels generated continuations and failures, and the resulting search trees are converted into supervised targets for three actions: continue, finish, and backtrack, with optional traces that summarize abandoned branches. We evaluate Pyligent on a hidden directed graph task designed to isolate delayed-failure recovery, and on structured reasoning domains with exact validators, including $4{\times}4$ Sudoku, Sudoku with reasoning traces, and Blocksworld. Compared with gold-only supervised fine-tuning, Pyligent improves solve rate by $72.7$ percentage points on hidden graphs, by $17$ and $18$ points on mixed and expert Sudoku, by $27$ and $14$ points on mixed and expert Sudoku with reasoning traces, and by $13$ points on Blocksworld. These results suggest that explicit failed-branch supervision can teach useful recovery behavior beyond imitation of polished solution chains.

Federated learning (FL) avoids explicit data exposure by keeping raw data on local clients, yet privacy risks remain in the training process and the learned model itself. Recently, centralized Taking Away Training Data (TATD) attacks have shown that malicious training could abuse the memorization capacity of deep models to store and later recover training data. However, this memorization-based threat has not been systematically studied under FL environments, where multi-client averaging could overwrite encoded training data. In this paper, we study a white-box TATD attack in which a malicious server selects n target clients from K participating clients and actively writes private training data into the global model during federated training. We propose FedCVESA, a federated variant of Correlation Value Encoding Attack (CVEA), by adding a Pearson-correlation regularizer to the loss function of target clients, so that private training data are gradually encoded into selected model parameters, referred to as carrier parameters. To reduce the overwriting of carrier parameters during server aggregation, we further propose segmented aggregation over dispersed carrier parameters, preserving selected carrier parameters while keeping standard averaging on the remaining parameters. Experiments on MNIST, Fashion-MNIST, and CIFAR-10 under Dirichlet non-IID partitions show that the proposed method can steal semantically meaningful private training images from the trained model while maintaining acceptable main-task utility in a controlled proof-of-concept setting. These results demonstrate that FL can become a parameter-level memorization channel for active TATD attack under the studied white-box malicious-server setting.

Large Language Models achieve strong code generation for high resource languages like Python and Java but suffer sharp performance drops on Low-Resource Programming Languages~(LRPLs) such as Julia. Improving Small Language Models~(SLMs) for these languages faces a trilemma: Supervised Fine-Tuning~(SFT) is bottlenecked by data scarcity, inference-time scaling is too expensive for deployment, and Reinforcement Learning from scratch yields near zero advantages. We propose a three-phase pipeline that resolves this trilemma by decoupling syntax acquisition from algorithmic reasoning. First, we \emph{left-shift} inference-time compute to an offline data synthesis engine that uses iterative compiler and test feedback to generate verified training examples. Second, we fine-tune an SLM on this synthetic, verified data to embed strong syntactic priors. Third, we apply Reinforcement Learning with Verifiable Reward~(RLVR) grounded by language-agnostic Input/Output tests, where the SFT prior constrains exploration away from syntax errors. Applied to Qwen3-8B, our pipeline improves pass@1 by up to +7.6 points on MultiPL-E and +14.2 points on the Agnostics LiveCodeBench for Julia compared to SOTA results. Furthermore, the pipeline only used $\frac{1}{3}$ data and $\frac{1}{6}$ cost over the previous state-of-the-art. We further demonstrate that the pipeline generalizes to Ballerina achieving 49.7% MultiPL-E Pass@1, a language with near-zero pretraining representation. Ablations confirm that both the SFT phase and execution-grounded rewards are necessary for stable training.

Implicit bias is usually explained as the preference of an optimization process for certain final solutions and their geometry. This view helps explain where a model finally stops. It gives less direct explanation of how this bias is formed during training. This paper proposes a training-time information allocation view. Under this view, optimization forms a writing pattern for error signals across parameter paths, coordinate channels, and sample regions. This paper builds a set of observable allocation diagnostics. These diagnostics include gradient demand, actual update injection, coordinate gain induced by exponential moving averages, channel-level update ratios, and sample-wise loss distributions. To separate training progress from internal allocation, this paper introduces a collapse--persistence analysis. Under matched training loss, if external loss statistics collapse but internal allocation ratios remain separated, then the factor changes the internal allocation of the training signal. Overall, this paper extends the analysis of implicit bias from final-solution geometry to training-time signal allocation. The main claim is that implicit bias is not only reflected by the final solution. It is also reflected by which parameter paths, coordinate channels, and sample regions receive the error signal first and more strongly during training. Based on this view, this paper places different training factors into a unified information-allocation diagnostic framework. The framework gives a mechanism-level explanation of training-time implicit bias. It also provides a basis for future optimization methods that control training progress and signal allocation separately.

Learned optimization aims to improve upon hand-designed optimizers (e.g., Adam and Muon) by meta-learning small neural network optimizers over a distribution of tasks. While recent work has greatly advanced the architectural design and inductive biases of learned optimizers (LOs), their meta-training remains biased toward short-unroll learning on particular tasks, resulting in redundant computation and leaving LOs often unable to compete with hand-designed optimizers. We introduce Efficient Long-hOrizon (ELO) learning, an efficient meta-training algorithm that (1) reallocates wasted meta-training compute to longer failure regimes, achieving efficient long-horizon learning, and (2) enforces decoupled progressive expert supervision, providing stable meta-learning signals that additionally improve the generalization of LOs. Our empirical study evaluates ELO for meta-training both element-wise and matrix-based LOs. Across downstream language modeling (GPT-2-124M/350M on FineWeb) and image classification (ViT-B/16, ResNet-50 on ImageNet-1K) tasks, ELO substantially improves the long-unroll performance and out-of-distribution generalization of the base LOs. In particular, ELO-Celo2 consistently outperforms well-tuned AdamW across all evaluated tasks, while remaining competitive with Muon on language modeling. \textit{Notably, all ELO baselines require less than 7 H100 GPU-hours for meta-training.}

Backdoor attacks severely threaten large-scale AI models. When model owners delegate training to external compute providers within a decentralized training paradigm, adversaries can craft stealthy, low-frequency triggers to inject malicious behavior while evading standard audits. Traditionally, detecting these attacks requires a full re-computation of the training steps--a prohibitive overhead that directly contradicts the owner's resource constraints. To address this, we investigate the resilience of continuous optimization dynamics under Byzantine perturbations, where adversaries are forced to compete against a continuous influx of honest updates. Under a threat model where an adversary compromises f out of n total trainers, we quantify the minimum auditing overhead required by the model owner to probabilistically bound the attack success rate. We formalize this injection-absorption dynamic as a Discrete-Time Markov Chain (DTMC). Using this framework, we prove that the success probability of any bounded adversary asymptotically collapses to zero under a defense strategy combining natural absorption, a randomized scheduler, and lazy verification oracle. Empirical results demonstrate significant backdoor suppression with zero utility degradation even when invoking the verification oracle on merely 10% of the total training steps. This approach yields a provably sound and computationally efficient defense for safety-critical AI.

Which properties of a partially trained network are causally portable to a different, independently trained network? Single-trajectory interventions show necessity within one run, not portability across runs. We introduce cross-trajectory chimera interventions: given two runs from different seeds, we split each weight vector into a norm and a unit direction, recombine one run's norm with the other's direction, and continue training. On two modular-arithmetic tasks that grok, the components dissociate. Direction carries a transferable, donor-specific circuit identity: implanting a donor's direction at the recipient's norm drives the run to the donor's circuit in 40/40 cases, while an angle-matched random control yields no shift. The transfer is threshold-like, and its location is predicted by the recipient's norm, separating perfectly by norm class over all 20 pairs (joint permutation probability 1.9e-4). Norm carries only a modest, distributed delay effect and no identity signal. An adaptive bisection procedure localizes the threshold to +/-1/64. Direction indexes which solution a trajectory approaches; norm governs how susceptible that identity is to being overwritten.

The training paradigm of large language models has shifted from traditional one-pass training to multi-epoch training, as reasonable reuse of limited high-quality data can improve both model performance and sample efficiency. Meanwhile, excessive repetition introduces the risk of overfitting and diminishing returns. Determining when and how to reuse data effectively thus emerges as a natural but under-explored question. Through a novel observation of model's "Memorization Window" signals derived from loss retention dynamics and downstream evaluation scores, we propose "Memorization-guided Data Reuse", a training paradigm that adaptively determines when and how data should be reused, enabling principled decisions on the number of training epochs and the scheduling of data replays. Our preliminary experiments reveal a consistent memorization-driven regime: performance continues to improve with repetition far beyond current practice (e.g., the commonly cited four-epoch limit). While a full scheduler remains future work, these insights provide a foundation for memorization-aware training schedules, helping to determine reuse budgets and move toward training LLMs smarter rather than longer with limited high-quality data.

Large generative models across text-to-text, text-to-image, and image-to-text modalities have been shown to pose significant privacy risks. One fundamental threat is membership inference attacks (MIA), which aim to determine whether a given data point was used in a model's training set. Although prior work has investigated MIAs against these three classes of generative models, existing approaches treat them in isolation and are not cross-applicable, thereby limiting their real-world utility. To address this limitation, we present the first comprehensive study of a unified membership inference framework that applies across text-to-text, text-to-image, and image-to-text modalities. Our approach is grounded in a key modality-agnostic observation: the output distribution of a generative model can approximate its training data distribution. Leveraging this property, we model the distributions of model-generated outputs and auxiliary non-member samples in a shared embedding space, and perform membership inference via likelihood ratio testing. We conduct extensive experiments in a strict black-box setting under both partial-knowledge and zero-knowledge threat models, and evaluate membership inference against both fine-tuning and pre-training data. Experimental results demonstrate our approach's superior performance in comparison to existing state-of-the-art methods, which are typically optimized for a single model class.

Orthogonalizing the mLSTM memory matrix at read time with five differentiable Newton-Schulz iterations improves noisy associative recall. We replicate this effect and investigate its mechanism. Training on MAD noisy recall exhibits a long chance-level plateau followed by a sharp increase in accuracy. The orthogonalized read improves conditioning during this plateau and can be removed after escape. Ablations support three findings. First, the benefit requires a self-consistent read and gradient: an exact recursive least-squares read (the Mesa layer) yields a similar benefit, while straight-through variants, delta-rule writes, frozen random keys, and Frobenius normalization show no improvement over baseline. Second, across a learning-rate x task-difficulty grid, orthogonalization multiplies escape hazard roughly six-fold, with no detectable dependence on difficulty, and widens the range of learning rates that produce successful runs. Third, adding orthogonalization at inference leaves chance-level failures unresolved, while removing it gradually after escape yields standard mLSTMs at near-perfect accuracy. Schedule changes alone recover much of the reported gain. A batch-size x learning-rate analysis separates the effects of per-step learning rate and gradient noise on escape hazard (elasticities +3.0 and -1.65, respectively), linking the original vocab-96 result to its large-batch training regime. Direct decoding of the memory state recovers roughly half of the associations in behaviorally failed models, indicating a readout-learning limitation despite substantial stored information. These results show that fixed-budget recall benchmarks are sensitive to trainability and provide a tractable setting for investigating abrupt behavioral transitions through measurements of internal representations.

Most LP-from-text benchmarks are static datasets of word problems written and labeled by hand. Once such a dataset is released, its size is fixed, its difficulty is fixed, and every problem can leak into the training data of future LLMs. We present \textbf{A$^{2}$utoLPBench}, a benchmark for testing LLM-driven agents on linear programming problems written in plain text. We first pick a feasible point and dual, then write down a problem for which that point is optimal and the objective value is known. The answer is known by construction, with no solver call and no human annotator. The evaluation environment bundles a reference solver-critic baseline and a Docker image whose usage instructions are written for an LLM-driven agent to read. With these in place, any agent can run the benchmark and get a calibrated score with one command. Because the benchmark is a generator rather than a fixed dataset, it has properties no fixed dataset can match: an unlimited supply of fresh problems, a difficulty knob set by $(n,m)$, ground-truth answers correct by construction, low LLM-side cost per problem relative to human authoring, repeatable scores across independent batches, and resistance to training-data leakage when fresh post-cutoff seed ranges are used.

In this paper, we describe NAVER LABS Europe's submission to the instruction-following speech processing short track at IWSLT 2026. We participate again in the constrained setting, developing systems capable of jointly performing ASR, ST, and SQA from English speech into Chinese, Italian, and German. Building on our previous submission, ranked first in last year's short track, we update our multi-stage training pipeline by replacing the speech projector with SpeechMapper, a method for learning a speech-to-LLM embedding projector using only ASR data. In addition, we introduce a synthetic SQA dataset, fakACL, composed of artificially generated scientific presentations. This dataset is built by prompting the LLM backbone, segmenting the generated talks, and synthesizing speech with SeamlessM4T-large-v2. The combination of an improved speech projection mechanism and domain-specific synthetic data allows our model to outperform last year's best short-track system, while being considerably more compact and relying on a weaker LLM backbone. This year's results place our system tied for first place in the overall short track ranking.

This paper showcases a memory-efficient training stack for Mixture-of-Experts (MoE) models. It is a training paradigm that combines and specializes various existing and novel parallelism techniques at different layers and stages of the Mixture-of-Experts (MoE) model training pipeline. It leverages these techniques to achieve maximal efficiency given the physical constraints of CPU, CPU memory, GPU HBM memory, and the CPU-GPU, GPU-GPU, and node-node communication bandwidth of the GPU cluster. It also contains a novel strategy for the optimizer step to achieve high throughput and memory efficiency, enabling practitioners to conduct lossless pre-training/fine-tuning of trillion-parameter scale models, at a million context length, with just under 12 8x H200 GPU nodes, with state-of-the-art throughput and memory efficiency. In our experiments, MoP delivers 4.7x--8.2x higher per-GPU throughput than a strongly-tuned FSDP2 baseline (with the gap widening at larger scale) and sustains training at context lengths up to 1M tokens, where the baseline runs out of memory beyond 64--128K.

Large Language Models (LLMs) have recently shown promise in molecular discovery, yet a gap remains between their probabilistic nature over discrete sequential tokens and the rigid topological constraints of chemical space. This raises the question of whether molecular LLMs can generalize beyond the local neighborhoods induced by their sequence-based representations. To systematically investigate this question, we introduce a Molecular Perturbation framework that generates syntax-valid structural variants of training molecules under controlled Graph Edit Distance (GED) to probe the manifold regularity of molecular LLMs. Our analysis shows that even a single edit can cause substantial performance drops on common molecular tasks, revealing a narrow local trust region and fragile sensitivity to structural changes. Since similar molecules tend to exhibit similar properties, In-Context Tuning (ICT), which anchors predictions on structurally similar molecules, offers a natural way to mitigate such fragility. Our experiments also examine whether ICT confers robustness under controlled structural perturbations, and the results suggest that it can partially expand the local trust region and offer a promising direction for stabilizing molecular LLMs against structural variation.

Foundation models are routinely released to the public, yet the data recipes used to train them -- such as domain mixture weights that determine how different sources are sampled -- are rarely disclosed. This creates an access asymmetry: researchers study the resulting models but lack visibility into the training distribution that produces them. Prior works for inferring training data, such as membership inference, detect at the level of individual samples and thus cannot characterize the global composition of the training corpus. We introduce WARP, a framework that recovers a fine-tuned model's training mixtures directly from its released weights. WARP interpolates between the base and fine-tuned models using model merging, generating pseudo-checkpoints that approximate the missing training trajectory and expose a geometric footprint of the training data in the weight space. From these simulated footprints, WARP extracts geometric features and maps them to domain proportions using either a parameter-free softmax readout or an MLP projector trained on synthetic mixtures. In controlled experiments with BERT and GPT-2, WARP recovers domain mixtures with an average MAE as low as 0.046 and 0.104 respectively, outperforming membership inference and a variant with access to the true training trajectory.

This work investigates the effect of batch sampling strategies during training for text-to-audio music generation under low-data and small-scale model settings. This paper describes our approach and findings for the ICME 2026 Grand Challenge on Academic Text-to-Music Generation. Training data are clustered using either text embeddings or audio embeddings, and samples with similar characteristics are grouped within the same mini-batch to mitigate gradient interference. The effects of modality and cluster granularity on clustering are analyzed. Results show that clustering based on text embeddings achieves better performance on objective evaluation metrics than clustering based on audio embeddings. In addition, different cluster granularity leads to different behaviors across evaluation criteria: a moderate number of clusters performs best on objective metrics, while a larger number of clusters tends to exhibit music with more coherent structure in listening tests.

We propose a scaling law that takes into account model size and training data while explicitly splitting the latter into training steps and batch size (called three-term law). Fitting the proposed law on a large set of training runs, we find that it correctly recovers the scaling of the optimal batch size. Moreover, because it makes use of training runs with suboptimal batch size, our proposed law can be robustly fit with a significantly smaller amount of training runs. We further show that the three-term law can be used to derive scaling laws for suboptimal batch sizes, and that it matches previous empirical findings related to the critical batch size.

GPU training jobs fail often, roughly two in five on large production clusters, yet the operator typically learns of a failure only by reconnecting hours later. Experiment trackers require editing the training script and maintaining a cloud connection; the scheduler's mail hook delivers a single status line with no cause and no logs. GPUAlert is a command-line wrapper that monitors any training command at the process boundary, and with no change to that command, emails a structured notification on completion carrying a classified failure cause, durable logs, and output artifacts. The tool is organized around three reliability primitives: a pre-launch log guarantee that establishes the durable destination before the child process can crash, notifier isolation that makes the wrapper's exit code a pure function of the child's status regardless of whether the email succeeds, and a non-silent artifact budget that bounds attachment size without ever dropping output silently. We release a labelled corpus of 474 GPU training logs across 15 failure classes and a reproducible evaluation harness. On the twelve hardware-reproduced classes, the ordered-rule classifier reaches 0.997 macro-F1, against 0.830 for unordered keyword matching and 0.133 for exit-code inspection. Wrapper overhead is a constant approximately 3ms per job; the pre-launch guarantee preserves a log where a shell redirect yields nothing; and across all 15 failure modes the wrapper returns the child's exit code unchanged even when the SMTP relay is unreachable.

A growing body of literature suggests that training data membership inference problems are fundamentally hard tasks in modern language modeling settings. We argue that output watermarking techniques are the right gadget to make training membership tests for generative models more tractable, based on prior results showing that language models exhibit residual watermark "radioactivity" under partially watermarked training datasets. We pit a watermark-based dataset inference approach head-to-head against traditional loss-based membership inference methods and show that watermarking can achieve comparable membership detection performance when subset exposure is high enough, under an alternate set of assumptions.

Semantic segmentation models struggle with data sparsity and rare or visually diverse regions, e.g., dense regions or small objects in aerial or autonomous mobility data. While synthetic augmentation is an appealing solution, directly generating new labeled data risks misalignment of labels and generated pixels. Existing solutions to this problem often rely on external models, or employ coarse heuristics such as indiscriminately augmenting all foreground objects or entire backgrounds, which wastes capacity on uninformative pixels. To address this, we propose an uncertainty-guided synthetic context augmentation strategy that strictly preserves label validity and efficiently maximizes pixel informativeness per synthetic sample - no external guardrails required. Using a baseline segmenter's predictive entropy, we identify uncertain semantic regions and inpaint only the complementary visual context. When fine-tuning the segmenter on this synthetic data, we compute the loss only over the original pixels, excluding inpainted regions. This focuses learning on the unmodified, uncertain regions while presenting them in novel contexts. We demonstrate substantial mIoU gains on Cityscapes, UAVID, and BDD100K with the largest gains on rare and difficult classes such as buses, trains, or (from the aerial perspective) cars. Our results demonstrate that uncertainty-guided context augmentation is a highly effective lever to improve segmentation performance on complex datasets, with code provided at https://github.com/XITASO/Preserve-the-Hard-Regenerate-the-Rest.

Text-rich image generation is one of the most challenging settings in image generation, since models must simultaneously produce visually realistic images and render legible, semantically aligned, and layout-consistent text. Existing data pipelines usually follow a static crawl-filter-freeze paradigm. They collect candidate samples, filter them once, and freeze the accepted data for training. However, rejected samples are usually discarded, although they often contain useful failure signals such as OCR errors and semantic mismatches. As a result, later construction rounds may repeat the same failure modes. To address these limitations, we propose DataEvolver, a self-evolving multi-agent framework for text-rich image data construction. DataEvolver treats data construction as feedback-driven construction policy evolution. A Retriever collects candidate samples, a Verifier assigns quality scores and rejection causes, a Critic summarizes round-level feedback into semantic feedback, and a Generator completes under-covered regions through targeted synthesis. The updated feedback memory then guides the next construction round. Experiments on text-rich image generation benchmarks show that DataEvolver produces more useful training data than fixed-dataset baselines under matched data budgets. At the 0.75M scale on PixArt-alpha, DataEvolver improves OCR-F1 over the strongest baseline by 85.3 percent on TextScenesHQ and 35.3 percent on LongTextBench. The improvements are consistent across both evaluated benchmarks and also transfer to Show-o2, indicating that the benefit of DataEvolver is not tied to a single downstream generator. These results suggest that rejected samples can provide actionable feedback for improving text-rich image data construction.

Enabling large language models to achieve stable self-improvement without external expert supervision remains a central challenge in complex reasoning tasks. Existing self-distillation and reinforcement learning methods lack explicit mechanisms for tracking problem-level learning progress and adapting optimization strategies accordingly. Consequently, training may over-optimize easy problems, receive weak supervision from hard problems, and fail to sufficiently explore borderline cases. To resolve these issues, we propose DRIFT, an online self-evolution policy optimization framework for large language models. DRIFT regulates the model's self-improvement process through the joint use of Difficulty Routing and Rhythm Gating. The former identifies the model's learning state at the problem level and dynamically allocates self-distillation and reinforcement learning signals, while the latter refines policy updates at the token level, concentrating exploration on critical reasoning positions. By further incorporating a success buffer and a two-stage curriculum learning strategy, DRIFT preserves high-quality historical experience while progressively guiding the model from reliable behavior acquisition toward stable policy evolution. Evaluated across five benchmarks and three model scales, DRIFT surpasses the peak performance of both GRPO and SDPO across all evaluated metrics. On the average score over the five benchmarks, DRIFT achieves 79.5$\%$, outperforming GRPO by 9.5$\%$ and SDPO by 7.5$\%$, establishing a new state-of-the-art result. Notably, on ToolUse, DRIFT reaches an accuracy of 79.2$\%$, improving over GRPO by 13.5$\%$ and SDPO by 10.7$\%$, setting a new state-of-the-art and substantially outperforming all concurrent methods.

Atomistic machine learning datasets are increasingly used for training: large immutable snapshots are read repeatedly, shuffled across epochs, staged across clusters' storage systems, and republished as reusable scientific artifacts. This workload differs from interactive scientific curation, where mutable records and ad hoc inspection are often more important than random indexed throughput. We present Atompack, an append-oriented storage format and distribution layer designed around a simple workload: training pipelines usually consume complete molecular records, while the order of records is randomized by the learning algorithm. Atompack appends records efficiently during dataset construction, then commits an immutable index and serves records through a memory-mapped read path optimized for training. We compare Atompack with HDF5, LMDB, and ASE baselines representing array stores, key-value records, serialized records, and object-oriented databases. The benchmarks measure sequential reads, shuffled reads, shared-filesystem behavior, write throughput, and artifact size. On a representative 64-atom workload, Atompack is 96x faster than ASE LMDB on shuffled training-style reads while producing artifacts about 79% smaller. The results indicate that serving complete molecule records, rather than field chunks or reconstructed objects, improves shuffled training throughput while keeping artifacts compact enough for public distribution.

We present the AI Training Manager, a bounded LLM-based metacognitive monitoring-and-control layer for machine-learning training. The manager asynchronously observes structured telemetry from an active training run, assesses the current training regime, and selects adaptive interventions through a constrained, deterministically verified action interface. We evaluate the approach on supervised and reinforcement-learning tasks. For supervised learning, we induce a multi-objective overfitting failure in a GPT-2-style model trained on TinyStories. The manager prevents the resulting late-run validation collapse, reducing final validation loss by 54.3% relative to the same stressed recipe. For reinforcement learning, we study robotic reaching under opposing multifactor stress regimes. In a conservative regime, the manager raises final deterministic safe success from 0.413 to 0.705. In an aggressive regime, it raises safe success from 0.121 to 0.764. The same manager instructions and action surface are used in both regimes. These results suggest that bounded LLM reasoning can provide a practical metacognitive control layer over ongoing learning processes.

Chart data extraction, which reverse-engineers data tables from chart images, is essential for reproducibility, analysis, retrieval, and redesign. Existing interactive tools are reliable but tedious, and mixed-initiative systems, while more efficient, lack generalizability. Recent multimodal large language models (MLLMs) offer a unified interface for chart interpretation, yet their ability to extract accurate data tables, especially without visible labels, remains unclear. We build a benchmark featuring diverse real-world charts without data labels to evaluate this capability. Results show that, while current MLLMs reliably reconstruct table structures, they struggle with precise value recovery. To address this, we revisit chart data extraction from a human-centered perspective and argue that extraction should follow a progressive learning process similar to how people read charts. Our training framework substantially improves numerical accuracy, achieving state-of-the-art performance with a 7B-parameter model. A user study further shows that our model effectively supports mixed-initiative workflows for reliable chart data extraction.

Symmetry provides a quantum neural network structure, but on its own it does not keep the network trainable once noise is present. We ask which physical quantity decides whether the gradients of an equivariant circuit survive decoherence, and we answer with a compact training law. Working with U(1)-equivariant brickwork circuits that conserve a charge, we find that two distinct effects govern a trainable gradient. Causality fixes where the gradient can live, confining it to the backward light cone of the readout inside the active charge sector. Coherence then determines how fast it decays through the contraction of the off-diagonal sector modes that the projected readout can actually observe. We prove a light-cone reduction that pins the noiseless gradient to the sector-restricted cone with a lower bound independent of the total qubit number, and we define a readout-visible aligned coherence rate as a Rayleigh quotient of the noise generator along the gradient-carrying mode. A perturbative open-system analysis turns this rate into a leading-order training law. Density-matrix simulations then confirm that the finite-noise degradation follows a single accumulated variable built from noise depth and coherence contraction, with a coefficient of determination of 0.979. The sharpest test comes from a correlated-dephasing channel that has a large worst-case rate but a near-zero aligned rate. The law predicts no gradient loss for this channel, and none is seen. Sector coherence outperforms every standard channel diagnostic we compare it against, and the analysis identifies readout-visible sector coherence as the quantity that links equivariant architecture, open-system dynamics and noisy trainability.

Recent large language models (LLMs) achieve strong performance on entity matching without requiring task-specific training data. However, applying these models to large sets of candidate pairs remains slow and costly. In contrast, entity matchers using traditional machine learning methods or small language models (SLMs), such as RoBERTa, offer much faster inference but require task-specific training data. This paper investigates whether the need to provide task-specific training data can be avoided by using knowledge-distillation workflows, in which an LLM serves as a teacher model to label training pairs that are subsequently used to train a smaller student model. We investigate knowledge distillation for entity matching along the following dimensions: pair-selection strategy, teacher model, label post-processing method, and student model. We evaluate the workflows using the Abt-Buy, Walmart-Amazon, WDC Products, DBLP-ACM, and DBLP-Scholar benchmarks, and compare the performance of student models trained with machine-labeled data to the performance of the same models trained using the benchmark training sets. Our experiments show that student models trained using the machine-labeled sets perform approximately on par with models trained on the benchmark training sets, with the remaining differences in both directions staying below two F1 points. Using GPT-5.2 to label the training sets for all five benchmarks costs US$28.31 to US$40.88, whereas manually labeling the same training sets is estimated to require 470 hours of work. At inference time, Ditto is 41.5 to 534 times faster than directly using an LLM to perform the matching tasks. These results indicate that current LLMs, when combined with a suitable pair-selection method, can substantially reduce or even eliminate the manual effort required to label use case-specific training data for entity matching.

Across social and online platforms, people are increasingly exposed to AI-generated images. As a consequence, the task of distinguishing AI-generated from authentic images is becoming a central challenge for information ecosystems. While humans perform better than chance, accuracy falls short of many operational needs. Initial evidence shows that visually oriented training can improve deepfake detection but does not improve participants' ability to identify real images as real. Here, we investigate the efficacy of a brief training intervention for intelligence analysts employed by the United States government in 2024. We conducted a counterbalanced within-subject randomized experiment in which we showed participants real and AI-generated images varying in pose complexity and scene context and asked them whether each image was real or AI-generated, both before and after an expert delivered a 30-minute training that pointed out patterns in seven real and 50 AI-generated images. We collected 2,544 image-level judgments from 32 intelligence analysts. We find training increased overall accuracy by 9 percentage points (95% CI: [2.7, 15.4]) from a baseline of 72%. We find the improvement is driven by a 14.2 percentage point increase in accuracy for real images (95% CI: [0.7, 27.7]). Through a careful experimental setup that curated matched pairs of real and AI-generated images across pose complexity categories, we reveal how these trainings influence people with different levels of digital forensics and generative AI experience and identify the kind of image-based content where this training intervention appears to be most effective. Ultimately, these results provide causal evidence that a brief, structured training can improve human judgment across a diverse array of real and AI-generated images, informing organizational responses to AI-generated visual misinformation.

The emergence of low-dimensional structures in the spectra of neural network weight matrices is a common empirical feature of trained models, but the dynamical origin of this phenomenon during learning remains an open problem. We formulate neural network training as the stochastic evolution of an initially random matrix ensemble, driven by stochastic gradient descent (SGD) updates that reshape the spectral bulk while amplifying signal strength. This induces a Baik-Ben Arous-Péché (BBP) transition during training, where isolated eigenvalues detach from the random bulk distribution, providing a dynamical framework for representation formation in high-dimensional learning dynamics. We demonstrate this in a solvable linear teacher-student model, where spectral evolution is analytically tractable and a phase diagram of trainability governed by the step size (or learning rate) and initial weight variance is obtained, and subsequently extend our formalism beyond the linear regime to nonlinear and stochastic settings. Numerical simulations in realistic settings support this picture, showing robust emergence of spectral alignment during training. Our results suggest that spectral analysis may provide a unified perspective of stochastic learning dynamics, linking trainability, optimisation hyperparameters, spectral phase transitions, and representation learning in neural networks.