Papers

Drug design couples property evaluation, conditional generation, structure-based design, and local optimization, yet machine learning systems typically address these capabilities with separate task-specific models. We introduce CAGenMol-2, a masked diffusion molecular language model that represents molecules, continuous scalar properties, and 3D protein pockets within a single wrapped sequence. Within this pretrained interface, downstream operations are selected by which sequence regions are observed or masked at inference, allowing one checkpoint to perform property prediction, property- and pocket-conditioned generation, and partial-constraint design without task-specific architectures or backbone fine-tuning. We further propose Adaptive Fragment Optimization (AdaFO), a gradient-free mask-and-refill search that turns the masked decoder into an iterative local molecular optimizer. On CrossDocked2020, AdaFO increases Success Rate from 30.2% to 70.8%, the best reported under this protocol, while largely preserving drug-likeness and diversity. Finally, scaffold-preserving directional editing and CRBN/VHL case studies demonstrate its use in compound design workflows spanning local molecular editing, structure-based prioritization, and downstream simulation-based screening.

We report a three-month autonomous research-agent program testing five solid-state-physics-inspired compression mappings on pretrained language models, with predictions committed to git before any pilot data and a 3-sigma gate deciding PASS or SHELVE. The common anchor – area-law / Kohn-nearsighted decay of the one-particle density matrix – has a distance face (P001 Wannier, P002 tight-binding) and a rank face (P003 DMRG-truncated MLPs, P005 Wilson-RG, P011 tensor-train embeddings). P005 was pre-empted at Phase 1; three of four Phase-3 pilots were falsified. On the attention face, GPT-2-medium attention-versus-distance is best fit by a stretched exponential in 12 of 16 median-layer heads once probe padding is excluded, and a tight-binding cutoff costs +96% perplexity (P002); on Pythia-160M the Wannier sparsity 0.054 +/- 0.004 is indistinguishable from PCA, random-Haar and identity baselines (P001). On the rank face, per-token tensor-train bond dimension does not track surprisal (r = 0.016 vs a pre-registered 0.65) and the format inflates rather than compresses (P011). P003 is mixed: its scaling claim shelved (r = -0.434), its MPO premise died at stage-0, and its cross-paper check, r = 0.523 as first written, collapses to 0.047 under the same correction, leaving both cross-paper checks null. The results invert the pre-registered prediction that most attention heads behave like Kohn-nearsighted insulators, pointing instead to critical, glassy or heavy-tailed regimes; the inversion is specific to the <= 350M scale tested, while the rank-face no-gain result held to 7-8B. We contribute the pre-registration + 3-sigma + cluster-framing + append-only-catalogue discipline – including why our own enforcement gate was designed but not deployed – four pre-registered negative results with full data release, and the inversion. The catalogue holds eighteen concluded studies, seventeen negative.

LLM pretraining corpora are normally cleaned by a stack of hand-written heuristics. A heuristic scraper extracts the main content from HTML, and dozens of rule-based filters then clean it, so corpus quality is capped by the coarseness and accuracy of the rules. In this work, we propose ReScraper, a unified language model of only 0.6B parameters that replaces this entire stack. To train ReScraper, we carefully curate supervised data from the outputs of three teacher models, so it learns to first extract the main content from raw data and then choose among four operations: keeping the page as extracted, editing out noisy lines and spans, deleting it entirely, or rewriting it when it is poorly written but informative. Based on the same crawled data pool, pretraining 400M, 1.4B, and 2.8B models on our curated data improves the DCLM Core score by a relative 3.8–4.7% over the strongest baseline at each scale, including the costly multi-agent curation. Our analyses show that each operation plays a distinct and complementary role, and that extracting and cleaning in one model outperforms a cascade of separate models. ReScraper also concentrates its operations on the pages that need them, raising the quality of poor pages the most while keeping the corpus diverse. These results demonstrate the feasibility and effectiveness of AI4AI for pretraining data curation, where a small learned model takes over an entire stage of the pipeline from hand-written heuristics. We open-source our code at https://github.com/cxcscmu/ReScraper

World models for manipulation are typically trained from video, yet the events that determine how manipulation unfolds, such as making and releasing contact, are difficult to observe visually and are often easier to sense through touch. We present the Dexterous Tactile World Model (DTWM), a video world model for future-frame prediction of egocentric manipulation from both observed video and tactile signals from a glove worn on each hand. We condition a pretrained video diffusion transformer on each hand's tactile signal through a zero-initialized residual at the corresponding hand location in the video tokens, while a causal mask prevents predicted frames from accessing future information. Compared with a vision-only model matched in architecture, parameters, and training, DTWM reduces the underestimation of hand motion from 23% to 9%, while reducing the perceptual error in the hand region by 7.4% across three training runs per model. The benefit also increases over the prediction horizon, with the improvement in the later predicted chunks being about 4.1x larger than in the first. DTWM also outperforms other visual-tactile world models under the same setting, and training with touch improves future-frame prediction even when no touch is available at inference. Ablations show that the model benefits from both the magnitude and spatial location of force: replacing the tactile signal with binary contact states, either per hand or per location, increases prediction error. The observed course of the force indicates whether the interaction will persist or change.

Imprecise annotations may support several plausible labelling distributions, yet learning methods often resolve this ambiguity into a single predictive distribution. This can obscure what the annotation evidence leaves unresolved. We introduce epistemic learning from credal supervision, a framework that uses convex sets of plausible labelling distributions, called credal sets, as supervision and learns sets of predictive distributions. We instantiate the framework with the pessimistic–optimistic credal classifier (POCC), which combines a shared backbone with two classification heads trained to minimise worst-case and best-case losses over the supervision sets. Their outputs define a predictive credal set whose spread provides an uncertainty score. We also show how credal labels can be obtained through a simple relaxation of existing probabilistic labels, reducing commitment to their precise probability assignments. This construction admits closed-form inner optimisation under cross-entropy loss, enabling efficient training. Assuming the supervision sets contain the true conditional label distributions, and other regularity assumptions, we establish a finite-sample generalisation bound for the averaged predictor with an explicit penalty for supervision imprecision. We evaluate POCC using human annotator disagreement and teacher predictions, alongside label smoothing as a controlled proxy for annotation imprecision. Across these settings, POCC achieves a favourable balance of predictive accuracy, calibration, and uncertainty-based selective classification versus competitive baselines.

High-dimensional Bayesian optimization (HDBO) seeks sample-efficient optimization when the number of variables is large relative to the evaluation budget. Recent LLM-based and agentic BO methods incorporate task knowledge and adapt search decisions during a run, but have primarily been evaluated on low- and moderate-dimensional problems. We ask whether this paradigm can transfer to the higher-dimensional regime. Our experiments show that these methods do not remain reliable in the high-dimensional regime, where the challenge is not only where to evaluate, but also which modeling assumption and search geometry to use when the objective's useful structure is unknown. We therefore introduce HERA, a Hypothesis- and Evidence-guided Research Agent that uses task context, optimization feedback, and structural diagnostics to revise search hypotheses, select and configure HDBO strategies, and determine their execution length. PRISM, its numerical optimization engine, generates and evaluates candidates sequentially within each search block, updating numerical models after each observation. HERA remains competitive with strong numerical HDBO baselines and outperforms the evaluated LLM-based and agentic methods on four metadata-free synthetic functions. Across eight real-world tasks, HERA achieves the best mean final objective among all evaluated systems on most benchmarks. Further analyses show that structural diagnostics change strategy use, metadata effects vary across tasks, and adaptive search blocks reduce inference cost.

Predicting molecular profiles from histopathology remains challenging because whole-slide images contain spatially organized, heterogeneous tissue patterns, while gene expression comprises thousands of correlated targets. We introduce MoSPR (Morpho-Spatial Program Regression), a linear framework that couples an adjacency-informed histology representation with a low-rank molecular basis. MoSPR clusters frozen patch embeddings into morphology microstates, aggregates their spatial adjacencies across the training cohort, and groups microstates with similar adjacency patterns into shared macrostates. Each slide is then represented by global morphology and macrostate-specific deviations, which are linearly mapped to coefficients of a training-derived low-rank gene-expression basis. Across three cancer cohorts from The Cancer Genome Atlas, MoSPR achieves the highest mean gene-expression prediction scores among all evaluated methods. Without pathway-level supervision, pathway scores derived from its predicted expression profiles rank first in eight of nine comparisons across three pathway collections. Ablation studies on the breast cancer cohort show complementary gains from adjacency-derived macrostate representation and low-rank molecular prediction. Moreover, with half of the training data on this cohort, MoSPR exceeds the full-data gene-prediction score of the strongest competing baseline. Finally, its linear formulation enables exact decomposition of each predicted expression profile into global and macrostate-specific molecular contributions, providing an interpretable link between spatially coherent macrostate regions and their associated molecular programs. Our code is available at https://github.com/Radisen-Panthera/MoSPR.

Self-evolving language models improve by generating tasks and learning from their own feedback, but adapting the task generator often requires a separate challenger-training loop. Can we generate tasks adapted to the current solver without explicitly training a challenger? We introduce Direct Self-Evolving Optimization (DEO), which replaces challenger parameter updates with solver-guided task sampling. The KL-regularized challenger objective defines an exponential tilt of a fixed base task distribution. DEO uses this distribution as a sampling target: a frozen LLM generates and mutates tasks, the solver scores them, and an approximate Metropolis selection rule refines the training pool. Only the solver is trained. Theoretically, for an idealized variant that samples exactly from the tilted distribution, and under regularity, local gradient-dominance, and initialization conditions, we show that DEO learns distributionally robust reasoning ability. In experiments, DEO achieves reasoning performance competitive with R-Zero while using over $50\%$ less wall-clock training time, and improves reasoning accuracy over a no-walk ablation. Replacing the task generator with a frozen API-only LLM further improves the local solver, illustrating a capability enabled by removing challenger training.

The projection of queries and keys are central to the attention mechanism in Transformer architectures. While they are mathematically symmetric, they play different roles in attention mechanisms. The question of whether there is an effect from their functional distinction on their geometric development in training remains unanswered. We investigate the problem through the training of small GPT-like Transformers on character-level WikiText-103 for three different depths (4, 6, and 8 layers), three types of initialization for queries and keys, and four random seeds, resulting in 36 runs and 54 trajectories of average layers across seeds. We track the effective dimensionality of those layers using participation ratios and discover that effective dimension of queries expand while keys shrink, and that $PR_Q - PR_K$ is positive in all trajectories studied. In connection to attention, the shrinking of keys leads to a narrower spectrum of $QK^\top$ and more peaked attention weights. In order to determine if this connection is causal or coincidental, we directly control the spectrum of keys during training across five seeds: restricting it to make it shrink sharpens the attention with high directional confidence, while keeping it constant to the level of initial dispersion makes attention softer. Additional token-level checkpoint analyses show that the monotonic paired-contrast trend is not universal across pretrained families, but survives as an early-training regime that later decays over a full pretraining run, and the link between interaction-rank geometry and attention entropy remains visible in several models.

BF16 is now standard in large-scale pretraining, including in fused attention kernels such as FlashAttention, and these kernels are widely trusted. When we used FlashAttention-3 to pretrain a 450M-parameter transformer on 50B tokens, however, we ran into a problem: training was healthy for 25B tokens, then the gradient norm grew a thousandfold and the loss ended 0.2 nats above FP32 attention, without a single NaN. Recomputing the attention backward of just two layers in FP32 removes almost all of the excess gradient. Part of the cause is known: a fused multiply-add in the forward softmax, so far treated as an extreme-input NaN case and never fixed in FlashAttention-3. Repairing it stops the blow-up, but the query gradient is still wrong by more than its own size, and training still drives attention logits to thousands of times their size under accurate gradients. The remaining error comes from a broken conservation law. The softmax score gradient sums to zero along every row, which makes the query gradient blind to where the keys sit as a group; rounding it to BF16 leaves a small nonzero sum that leaks the mean key into the gradient, and the leak grows exactly as late training makes keys large and attention sharp. We introduce GProj (gauge projection), which restores the zero sum after the cast with two rank-one corrections per row. It cuts the remaining median query/key gradient errors from 219%/13% to 0.34%/0.37%, on par with FP32 attention, for 4.7% more time per training step. In matched from-scratch runs it trains to the same loss as FP32 attention, while FlashAttention-3 and key smoothing both destabilize.

Error-correcting codes (ECCs) are essential across diverse applications, from wireless communications and storage to quantum computing, yet each application imposes distinct design requirements on the parity-check matrix (PCM). To address these on-demand requirements in a unified framework, we propose ZeroCode, a reinforcement learning (RL)-based approach that constructs PCMs sequentially from the all-zero matrix. ZeroCode formulates construction as a discrete sequential decision-making problem and uses proximal policy optimization with action masking to select valid edges. ZeroCode achieves a gain of approximately 1 dB over the prior RL-based construction method at a bit error rate (BER) of $10^{-4}$ for the (32,16) code and outperforms existing genetic, differentiable, and classical code-design methods in our experiments. Beyond optimizing decoding performance, the masking mechanism allows on-demand structural constraints, such as a maximum degree, 4-cycle-free structure, and quasi-cyclic structure, to be flexibly incorporated. Moreover, a single policy rollout yields a library of PCMs with varying edge counts, offering trade-offs between decoding performance and complexity without retraining. Overall, ZeroCode addresses diverse code-design requirements within a unified framework, providing solutions with optimized decoding performance under given constraints.

We propose a novel optimizer for low-rank adaptation (LoRA) that explicitly incorporates the gauge symmetry of low-rank factorization. Our optimizer extends recent matrix optimizers for full-parameter training to the manifold of fixed-rank matrices by interpreting them as normalization under a rotated basis. We show how rotation and normalization can be integrated with the fixed-rank manifold efficiently. Our optimizer converges faster to lower held-out loss and achieves better or comparable downstream performance on both supervised finetuning and reinforcement learning tasks.

Agentic benchmarks guide model selection and training. Yet an agent can pass a task without demonstrating the intended capability. Such outcomes constitute unearned passes; their proportion among all passes defines the integrity gap. As agents improve, benchmark surfaces that once seemed harmless can become exploitable, making benchmark validity an ongoing maintenance problem. We introduce a process-verification framework that audits passing trajectories, distinguishes evidenced reward hacking from verifier weakness, and localizes exploitable surfaces for repair. Across 3,810 passing trajectories from 29 model-benchmark cohorts, confirmed violations often increase with model generation but not monotonically. On SWEBench Pro V1.0, confirmed violation rates rise from 24% to 73% between Opus 4.7 and Fable 5 on matched tasks; later cohorts fall to 11% for Fable 5.1 and 0% for GPT-6 Astra. These comparisons are descriptive: configurations were not normalized, and the latest models also pass fewer exploitable tasks. Violations concentrate around a small set of recurring surfaces, especially unintended access to reference solutions through git history. Three repair case studies across two benchmarks show why blocking a recorded exploit is insufficient: the same protected information can remain accessible through another route. Therefore, we combine minimal patches with exploit replay and fresh agent evaluation, auditing new passes under the original standard. No evaluated attempt against the final patches reached the protected channel, and every post-patch pass was judged legitimate. Benchmark integrity requires ongoing maintenance: audit passing behavior, repair the enabling surface, and re-evaluate both exploit access and legitimate solvability.

General-purpose vision-language models offer a promising way to zero-shot robot control: \gptastra{} excels at open-ended and language- or image-conditioned manipulation but remains substantially weaker on high-precision and long-horizon tasks. We introduce RoboICL, an in-context robot-control framework that narrows these gaps without robot-specific parameter updates or a learned VLA. RoboICL separates demonstration context, which provides recorded examples when available, from interaction memory, which accumulates the model's own actions and observed outcomes. Both use a shared observation–action–receipt–observation grammar. To preserve experience across task stages, RoboICL combines sampled demonstration blocks with bounded anchored memory. Fixed anchors keep earlier rollout interactions available for in-context learning, while the latest interaction supports immediate error correction. Across 30 RoboDojo tasks, using zero shot for Open and one demonstration elsewhere, RoboICL improves on official zero-shot \gptastra{} by 20–27 progress-score points in every category. It leads the leaderboard baselines on Memory and Open, achieves comparable performance to the strongest Precision baseline, and remains competitive on Long-Horizon. Its 30-task Overall score is 50.64, versus 33.68 for the strongest baseline. On a separate ten-task subset, RoboICL scores 60.60, within 2.00 points of the $\pi_{0.5}$ + \gptastra{} hybrid approach. On three real-robot tasks, mean progress rises from 14.45 at zero shot to 63.33 at one shot and 78.89 at three shots. On two development tasks, optional Jev-gated action reuse reduces \gptastra{} calls by 33–48%. Code is available at \href{https://github.com/Mosi-AI/RoboICL}{https://github.com/Mosi-AI/RoboICL}.

Repeatedly training language models on their own generated data may create a synthetic-data feedback loop in which errors and distributional biases are reintroduced into subsequent training datasets. This paper studies that process in biomedical question answering (QA) using PubMedQA and two Qwen2.5 model sizes, 0.5B and 3B parameters. The study compares a recursive synthetic-data condition, in which generation G(k+1) is trained on answers produced by G(k), against a Human-Control condition that repeatedly uses the original human training data. The study evaluates across four generations from G0-G3 with two random seeds (42 and 123) and a fixed evaluation set of 1,000 expert-labeled samples. The evaluation includes disease and chemical entity F1, context-supported rate, lexical and semantic similarity, answer length, repetition rate, and other evaluation metrics. The Recursive condition for both model sizes and both seeds showed larger declines than the Human-Control condition in disease entity F1, chemical entity F1, context-supported rate, ROUGE-L, and cosine similarity. Under the fixed no-repeat 3-gram decoding constraint, the main observed behavioral change was increased answer length, while the measured 3-gram repetition rate did not increase. The magnitude of the difference-in-change was larger for the 3B model than for the 0.5B model. This difference was particularly apparent in disease F1, context-supported rate, cosine similarity, and answer length. These results show domain-specific changes associated with using recursive synthetic-data training in biomedical QA, but do not establish clinical hallucination rates or universal model collapse.

Pretrained autoregressive Transformers represent a large sunk investment in compute. Existing conversion methods reuse that investment by changing either the architecture (attention to recurrence) or the objective (next-token prediction to denoising), never both. We convert Qwen3 teachers at 1.7B and 8B into attention-free, bidirectional, gated-delta-rule diffusion students in three stages, so that each capability can be traced to the stage that kept or lost it. Language modeling transfers only partially and in-distribution; in-context retrieval does not transfer. On a multi-query recall probe where the teachers score 0.34-0.58, both converted students score 0.000, and diffusion pretraining alone does not restore retrieval. A retrieval curriculum in the final stage, which gradually lengthens the gap between a key-value table and the queries that address it, restores it only stochastically: on a fixed schedule, one seed in three learns to retrieve. Advancing the gap only while a running accuracy estimate stays above a threshold works for all three of those seeds, holds on real text, and carries unchanged to 8B, where two of three seeds succeed. The third had not learned within its fixed 16k-step budget: retrieval switches on abruptly at a seed-dependent step (6.5k and 11k in the other two), so a fixed budget can cut a late run off. One boundary survives every intervention: every model that learns retrieval scores 0.000 on tokens that never appeared in a retrieval episode, and an arm that resamples the key and value tokens every batch shows this is a coverage limit, not memorization of particular bindings. Separately, we convert a 7B code model into a 3:1 recurrent-attention block-diffusion hybrid over 85k steps and report two negative training results.

Autoregressive neural emulators can drift or diverge over long rollouts despite accurate short-term predictions. We introduce Cascaded Emulation (CasEm), a one-way rollout architecture that augments an existing full-state backbone with an independently evolving model of physically specified aggregates. Its forecasts guide corrections to full-state predictions, without feedback from the backbone to the aggregate model. Effective guidance requires aggregates that cover substantial backbone error, remain accurately predictable, and support useful full-state corrections. We derive a finite-horizon error bound that clarifies these three factors and use empirical diagnostics to guide subsystem selection. Across four ODE/PDE benchmarks, CasEm reduces long-horizon rollout errors across diverse backbones and suppresses the trend toward error divergence in both diffusion tasks using Fourier neural operator backbones. In global climate emulation, CasEm with a regional total-water subsystem reduces 10-year full-state time-mean error by 66.6% and 46.3% for frozen ACE and Spherical DYffusion backbones, respectively, while adding less than 3% to inference time.

LLM agents increasingly take privileged, often irreversible structured actions, such as paying an invoice. They assemble each action from action-critical fields in documents and tool outputs that an adversary can corrupt, and indirect prompt injection can drive the model itself to extract attacker-chosen values. Current defenses gate on a source's trust label or certify free-text answer quality. None certifies the integrity of a coupled, policy-bound structured action under a corruption budget that accounts for shared upstream sources. We characterize when such an action is safely certifiable and give the maximally live safe certifier. It admits an action only when each field clears the rule its evidence structure supports: a bounded corruption radius over corruption-distinct evidence classes, counted by a minimum hitting set so that re-publishing or laundered copies cannot manufacture a quorum, deterministic reconciliation for complementary fields, and a trusted anchor where the evidence leaves a field single-sourced. We formalize two robustness notions, validate each mechanism by ablation, and measure how often the multi-source precondition holds on sanctions designations (70,966 entities) and software supply-chain provenance (450 packages). Under upper-bound proxies, genuine corroboration is a minority phenomenon in both, and naive attestation counting overstates it, since witnesses that look independent collapse to two corruption-distinct domains once shared origin is counted. Across five current models in a real agent loop, a realistic injection fools every model but one and a naive agent then executes the fraudulent action on most attacks. The certifier admits no unsafe action and recovers the correct value where corroboration permits, while action-gating and provenance baselines are broken in every world of our harness by some attack in its space.

AI agents require memory that remains consistent as conversational evidence changes. We present Stashbird, an episode-grounded memory system that uses provenance to reconcile derived graph and retrieval state during updates and deletion. Repeated evidence extends support, corrections supersede prior active relations, and deletion removes unsupported state while preserving relations with remaining support. Under fixed semantic decisions, reconciliation produces correct final states in all 20 update sequences, compared with eight without reconciliation, and satisfies the defined checks across 263 deletion interventions. Across four long-term memory benchmarks, Stashbird achieves competitive question-answering accuracy and strong model efficiency. On LoCoMo, it uses 76.4x fewer ingestion prompt tokens than Graphiti and 8.1x fewer retrieval prompt tokens than reproduced Hindsight

Evaluating statistical discovery by large language model (LLM) agents requires verifiable analytical ground truth. Establishing such ground truth for real-world datasets is costly, and prior knowledge of public datasets can influence agent responses. We introduce SLEUTHBENCH, a benchmark that addresses both problems by injecting controlled data-quality problems and feature effects into public tabular datasets: the injected pattern determines the answer, so reference answers are computed automatically and memorized knowledge of the original table is insufficient, while the table keeps its background structure. The injected patterns are modeled on phenomena reported in real data analyses. The benchmark defines 17 question templates in two families: data-quality questions and feature-contribution questions. We evaluate six state-of-the-art LLMs that analyze the data using a Python coding tool, on data-science and business phrasings of 70 validated dataset-template combinations, yielding 1680 graded responses in total. The models detect data-quality problems reliably (83.8% accuracy) but recover feature contributions poorly (41.9%). Finding how features shape the target requires searching over both candidate variables and analytical procedures. To address this issue, we propose the Empirical Layer, a set of precomputed statistical artifacts comprising summaries, fitted feature and interaction effects, and dataset descriptions, which exposes candidate patterns for direct inspection. Access to these artifacts raises feature-contribution accuracy from 41.9% to 68.0%.

Looped language models separate computational depth from parameter count by repeatedly applying the same transformer block. Adapting these models requires a shared update that remains effective as hidden states evolve throughout the recurrent computation. Our empirical analysis reveals a pronounced late-loop bias in standard low-rank adaptation (LoRA): the shared update provides limited adaptation at early loop positions. This imbalance motivates training shared updates under varying combinations of their applications. Randomly omitting adapter applications alone, however, does not improve task performance; it reduces expected update strength during training while leaving inference unchanged. We introduce Loop Dropout, which couples stochastic masking of adapter applications with inverse-survival rescaling to preserve expected update strength and promote effective adaptation across loops. Extensive experiments demonstrate improved mathematical reasoning across model sizes, adapter ranks and training recipes, with benefits extending to general instruction tuning and code generation. Loop Dropout outperforms existing LoRA variants and adapter regularizers, while further analysis shows stronger early-loop adaptation. Every backbone loop remains active, and inference applies the adapter at all loops using standard LoRA without additional trainable parameters or inference computation. Code is available at https://github.com/NUS-HPC-AI-Lab/loop-dropout .

Evaluating how LLM agents recover from mid-task failures is central to deploying reliable agentic systems. Existing checkpoint-based benchmarks measure recovery by comparing which action is selected as best across independent runs, a quantity known as set agreement. However, set agreement is a purely ordinal measure that records which action wins without reflecting the absolute level of performance. When all actions fail, they tie at zero reward, and independent runs produce the same tied set with high probability, creating an illusion of stability that masks near-zero recovery success. We formalize this limitation through a set-path symmetry result, proving that for equal-cost Bernoulli actions the success probabilities (0.9, 0.8) and (0.2, 0.1) yield identical best-action-set distributions at every sample size. No procedure based solely on which action wins can distinguish these two regimes. We further prove that certifying exact population ties is impossible in finite time, and that the assignment of outcomes to checkpoints carries information beyond marginal outcome distributions. The pooled success probability is the missing scalar that resolves the ordinal ambiguity. Experiments on 864 frozen RecoveryBench episodes and two planning cohorts totaling 3,456 responses confirm the theoretical predictions. Agreement and held-out quality can move in opposite directions, and permuting checkpoint-to-action bindings changes 8 to 13 percent of cell-level conclusions. Based on these findings, we propose reporting four diagnostic quantities (agreement, all-zero fraction, held-out success, and pooled success) that expose this failure mode with no additional data collection.

Mixture-of-Depths (MoD) enables conditional computation across Transformer depth by routing only a subset of tokens through selected layers, but its original one-sparse–one-dense alternation tightly couples total capacity to active capacity and limits sparse-depth scaling. We introduce X-MoD, a scalable sparse-depth architecture that decouples token sparsity from anchor stride, allowing total parameter count to grow while keeping active-equivalent capacity nearly fixed. To make deep sparse routing trainable, X-MoD combines dense anchors with variance-scaled layer-wise gating and depth-wise token balancing. To make this regime analyzable and usable, we formulate sparse-depth routing as a conditional architecture-design problem: given compute, context length, and active-equivalent backbone size, how should the routing configuration be chosen? We develop a practical scaling-law framework by fitting X-MoD relative to FLOP-matched dense baselines, yielding an interpretable law that decomposes performance into sparse-capacity gain, sparse-context correction, and anchor-stride interaction. The law predicts validation loss across routing configurations and reveals how context length, model scale, and anchor stride shape sparse-depth performance. We validate the architecture and law through pretraining sweeps, held-out scaling-law prediction, ablations, downstream evaluations, and comparisons with Dense, MoD, and representative MoE baselines.

Functional data are intrinsically infinite-dimensional, and often exhibit phase variation, where corresponding events occur at different times across observations. Existing linear dimension reduction methods struggle with nonlinear amplitude variation, while functional autoencoders without an explicit warp entangle temporal misalignment with shape. We propose the Amplitude–Phase Functional Autoencoders (AP-FAE), an unsupervised framework for functional data that spans both univariate and multivariate cases, with emphasis on the multivariate setting, and factorizes the latent space into separate amplitude and phase embeddings derived from all channels. A smooth functional decoder reconstructs channel-specific amplitude functions in canonical time, and a shared monotone, endpoint-preserving warp captures phase variation. We prove a bound linking amplitude recovery to registration, reconstruction, and noise errors, and validate it numerically. Across synthetic data and six real-world benchmarks, AP-FAE outperforms state-of-the-art baselines on most clustering and alignment metrics and on all reconstruction metrics. Clustering with amplitude embeddings alone consistently surpasses joint amplitude–phase clustering, confirming the benefit of explicit disentanglement. Code is available at https://anonymous.4open.science/r/APFAE-418C/}{https://anonymous.4open.science/r/APFAE-418C/.

Optimization modeling is central to many decision-making scenarios, but traditionally requires extensive domain expertise. While Large Language Models (LLMs) have shown promise in automating this process, current training paradigms mainly rely on human-annotated or teacher-generated datasets. This dependence introduces a Generalization Ceiling, where models overfit to narrow data distributions, and Capability Anchoring, where models' reasoning is bounded by annotator proficiency and teacher model capability. In response, we propose OPT-Zero, the first fully self-play training framework for optimization modeling that requires zero external training data. OPT-Zero employs a single LLM in a dual-role closed loop: a Proposer that synthesizes increasingly challenging optimization problems alongside their mathematical formulations and solving code, and a Solver that attempts to resolve the problems given only natural-language problem descriptions. Grounded in execution feedback from external optimization solvers, we alternately train both roles using reinforcement learning. This process fosters an auto-curriculum in which the Proposer and Solver co-evolve: generating harder valid problems by the Proposer seamlessly enhances the structural reasoning ability of the Solver. Extensive results indicate that with zero curated data, OPT-Zero matches state-of-the-art data-dependent methods while exhibiting substantially stronger generalizability, establishing self-play training as a highly scalable paradigm for advancing LLM reasoning in modeling and solving optimization problems.

Language-model judges compare agent upgrades with their predecessors, but a fixed judge can make version-dependent mistakes. We analyze 35 public coding-agent submissions (20 prespecified version pairs on 250 SWE-bench Verified issues), two customer-service agents (155 tau-bench tasks), and 1,106 expert-labeled AgentRewardBench trajectories. An upstream outage left three judges for the primary SWE-bench analysis (8,743 aligned cells); the fourth is descriptive. All three coding-agent judges and all four tau-bench judges reject task-conditioned error invariance after multiplicity adjustment. On SWE-bench, 32 of 60 judge-by-pair units have a detectable differential comparison component; eight judge-only intervals declare upgrades that execution-based intervals cannot establish, despite rank correlations of 0.71-0.79. In tau-bench, one judge reverses a nine-point reference-reward gap by penalizing a procedural habit the reward ignores. False acceptance of failed coding patches rises with agent capability conditional on task and execution outcome, while a task-solvability prediction reverses sign across domains. A separately fixed post-submission OpenHands follow-up on the same 250 issues (eight configurations, 1,981 three-judge cells) reproduces the capability/false-acceptance association (mean Spearman +0.944, exact p=0.000099) and decreasing Youden contrast (mean -0.937, p=0.000397); this is observational, not a new-task replication. Transporting old-version calibration raises SWE-bench comparison error from 3.8 to 19.5 points, with 24.6% undefined bootstrap ratios. A paired audit saves only 5% in interval width at 80 labeled tasks. A randomized self-report test is negative (three adjusted p-values=1.0). These results favor paired audits of current outputs over judge-only release decisions or transported calibration; independent human patch review remains pending.

Transactional basket data can reveal associations among items, but observed co-occurrence conflates item-specific relations with basket-size structure and unmodeled higher-order dependence. We introduce Cardinality-Stratified Interaction Decomposition (CSID), an interpretable framework that decomposes log-odds contrasts stratified by the number of remaining items into item-set-specific and cardinality-common components, without fitting a global joint distribution. CSID uses an information-weighted, gauge-constrained ridge projection to estimate pair and triple components and to diagnose higher-order contributions to pairwise structure. CSID is designed primarily for interpretable decomposition of association structure rather than for full-distribution prediction. In a simulation with zero pair effects, increasingly strong small-basket cardinality potentials drive ordinary Ising couplings spuriously negative, whereas CSID pair estimates remain centered near zero. Detection power rises with the magnitude of planted triple effects, and local deprojection reduces pair-coefficient RMSE from 0.244 to 0.073. Across three grocery datasets, high-information triple components are reproducible over time. In the matched cross-period partial-transfer evaluation, transferred CSID triple components show closer agreement with later-period stratified contrasts than the nodewise-symmetrized cardinality-aware higher-order pseudolikelihood comparator, with gains in weighted Lin's concordance correlation of 0.038–0.122. These results support CSID as an exploratory and interpretable decomposition framework for pairwise and higher-order association structure in transactional data.

Formulaic alpha discovery is a core challenge in quantitative trading, as identifying alphas that work well together remains difficult. Recent reinforcement learning (RL) methods formulate this task as a Markov decision process (MDP), but two important issues remain unresolved. First, as the alpha pool evolves, the reward function changes accordingly, making the MDP inherently non-stationary. Second, most existing methods optimize a single objective, typically predictive power, while ignoring other important properties of a high-quality alpha pool. Motivated by these challenges, we propose AlphaPareto, an RL method for formulaic alpha discovery. To address non-stationarity, AlphaPareto augments the state to include both the alpha under construction and the current alpha pool, and applies a large language model (LLM) to encode the pool. This design allows the agent to adapt to the evolving search environment. To overcome the limitation of single-objective reward design, AlphaPareto replaces the scalar reward with a multi-objective vector-valued reward that simultaneously captures predictive power, temporal stability, perturbation robustness, and diversity, and optimizes these objectives through a Pareto-regularized learning procedure. Empirical applications to real-world datasets show that our AlphaPareto method outperforms its competitors.

Weight compression helps large neural networks fit deployment memory budgets, but common fixed-width formats offer only coarse storage choices. Entropy coding supports finer rates, yet the achieved size depends on the quantized weight distribution and coding overhead. Exploiting this flexibility requires accurate rate selection and efficient weight reconstruction for inference. We present EntroPack, an entropy-coded weight compressor that supports arbitrary target bitrates without activation calibration or fine-tuning. It combines row-normalized $E_8$ lattice quantization with a conditional probability model of lattice coordinates. Sampled storage estimates select the quantization resolution without repeated full-stream encoding. The final coordinates are entropy-coded in independently decodable tiles, enabling fast, fused symbol decoding and numerical weight reconstruction on the GPU. EntroPack supports floating-point and integer weight containers, such as BF16, FP16, FP8, and INT8, with storage bitrate controlled independently of numerical precision. Online decoding adds latency that grows with weight count, making the method well suited to compute-intensive workloads such as diffusion denoising and Transformer prefill. Experiments demonstrate fast encoding and modest inference overhead in these settings. When compressing the linear-layer weights of the image generator Z-Image-Turbo, EntroPack achieves substantially lower weight and denoiser output errors than fixed-width formats at comparable storage rates, with modest denoising-step overhead. Targeting 4 bits per parameter, it achieves lower weight and denoiser output errors than NF4, including about 24% lower relative $L_2$ weight error, with less storage. Source code is available at https://github.com/modelscope/entropack.

Retrieval-augmented generation systems are extensively instrumented with metrics, benchmarks, traces, and automated judges, but these tools do not decide whether a proposed policy change is safe to release. We present RAGWarrant, an open-source promotion-control framework that treats deployment as a constrained evidence decision rather than a leaderboard choice. RAGWarrant normalizes evaluator outputs and operational telemetry, applies predeclared quality and hard-risk gates, assigns evidence-class claim ceilings, preserves negative outcomes, and emits auditable PROMOTE, BLOCK, REJECT, or INCONCLUSIVE decisions. We evaluate the framework across T2-RAGBench, MultiHop-RAG, CRAG, HotpotQA, synthetic reproduction, and bounded local generative experiments. On HotpotQA, operational savings were blocked because answer quality fell beyond the declared margin. A bounded CRAG study selected a lower-cost quality-tied policy, but related generative gains were unstable and a held-out guardrail failed closed. We claim an auditable promotion-control abstraction, not optimizer superiority, human validation, or production readiness. The tagged artifact reproduces from a fresh clone, runs as a hardened Docker job, accepts external evaluator exports, and verifies artifact integrity.