#01Oct 26, 2026
cs.LG
Audit Where Hacks Pay: Budgeted Strong-Verifier Audits Against Reward Hacking in GRPO
Shurun Li, Michael Hua Wang, Haibo Zeng and 1 more
Reinforcement learning with verifiable rewards trains on cheap checks, such as unit tests, that a policy can learn to game, while a stronger check is too expensive to run on every rollout. We study how to spend a limited strong-verifier budget in GRPO, separating which weakly passing rollouts to audit from what reward a caught hack receives. At a fixed policy state with centered advantages, the expected update depends on a design only through each rollout's marginal audit probability and correction; reflecting a caught hack's reward about the failure reward negates the advantages of an all-hack group; and auditing whole groups with the fewest weak passes first (low-pass budgeted audits, LB) minimizes the advantage mass left on hacks among pass-count designs with a fixed correction when a label's value falls with the pass count. We tested this analysis in actual LLM training: from each run's starting state, the sign of this mass on a frozen panel predicted whether hacking would be suppressed in 33 of 39 runs, including 13 of 16 preregistered ones. In LeetCode (Wong et al., 2026) and Countdown-Code (Khalifa et al., 2026), LB drives hacking to near zero, while random audits with the same penalty and budget do not and use more labels. We also find that the randomness of the update affects training: in Countdown-Code, two unbiased corrections, per-item reward-level IPW and a group-level advantage correction, bring hacking down too but end 6.1-11.8 and 8.0-11.5 points lower in held-out accuracy than LB at the same budget. A full audit, which the group-level correction reproduces in expectation, stays within about two points of LB in three seeds, while LB uses 28% of its labels, and allocating the same correction by pass count recovers most of the loss (one seed). Training outcomes thus depend not only on the expected update but also on the randomness of its estimate.
#02Oct 26, 2026
cs.LG
Quantitative Generalization Bounds for Multi-Layer Transformers: A Few Long Strings can Beat Many Short Ones
Marius Henry, Michael Rizvi-Martel, Michael Hahn
Transformer-based language models exhibit an astonishing capacity to generalize to longer and more difficult problem instances than those seen in training. However, theoretical understanding of this phenomenon still lags behind. Unlike most machine learning models, the generalisation abilities of transformers inherently require considering two axes. On one hand, one must examine in-distribution statistical generalisation, which concerns the ability of a model to broaden a function from training samples to novel samples of the same distribution. On the other hand, given the sequential nature of language data, one must also consider length generalisation, which reflects whether a model trained on inputs of a given length can extrapolate to longer ones.These two notions have mostly been studied separately, leaving open how they interact. In this work, we give a joint treatment of both axes for fixed-precision transformers. We give generalisation guarantees in terms of the parameter sizes of the involved models, as well as the quantization space and the alphabet size that supply the sequences. Finally, we illustrate how the requirements in the sample count and training length trade off. In particular, we find that increasing training length can be exponentially more efficient than increasing the sample size. Experiments support predictions of the theory.
#03Oct 26, 2026
cs.AI
Beyond Physical Illness: Can Clinical Narratives Reveal Emerging Mental Health Risks in Chronically Ill Patients?
Sudeshna Jana, Siddhartha Laha, Manjira Sinha and 2 more
Mental-health risks such as depression, dementia, and suicidal ideation can emerge during hospitalization among chronically ill patients and may worsen without timely recognition and intervention. However, most existing mental health studies focus on patients with established psychiatric conditions rather than patients hospitalized primarily for physical illness. In this work, we investigate whether latent mental-health risk in chronically ill patients can be inferred from longitudinal clinical narratives. We construct a weakly supervised cohort around 4,000 patients from MIMIC-III using mental-health evidence extracted from discharge summaries. A hierarchical state-space framework derives mental-health representations from day-wise clinical narratives and models their temporal evolution for risk estimation throughout hospitalization. Preliminary results demonstrate improved binary risk prediction using the proposed representations and show distinct model derived risk trajectories between risk and no-risk cohorts, providing a foundation for studying emerging mental-health risks in chronically ill patients.
#04Oct 26, 2026
cs.LG
Detecting bus delays and service anomalies in Niterói with autoencoder latent feature representations
Matheus Germano, Claudio E. C. Campelo
Bus riders judge service by two failures: single trips that run late, and days on which the whole system misbehaves. We ask whether one unsupervised representation of raw GPS telemetry can serve both diagnoses on the NetMob 2026 Niterói bus dataset. An autoencoder compresses each trip into a 64-dimensional latent vector; the frozen encoder feeds supervised delay classifiers, and the reconstruction error serves as an anomaly score. Two findings are negative. First, the latent representation does not improve delay detection: gradient boosting on per-trip summary features reaches $0.846 \pm 0.013$ ROC-AUC against $0.838 \pm 0.016$ for the best latent setup, and the gap becomes significant once the preprocessing filters are relaxed. Second, two of the three external checks of the anomaly score were circular. Retraining without rainfall removes the rain association ($5.4\%$ against $5.0\%$ flagged, $p = 0.84$) and an apparent citywide disruption on 23 March. What survives is a day-level signal stable across three autoencoders (Spearman $ρ\geq 0.96$), driven mostly by day type, with degraded-telemetry days still separable within weekday and weekend strata. The alarm threshold limits detection more than the score does: a pooled percentile cut flags $5.7\%$ of trips on the worst degraded day, while re-scaling within each bus line flags $29.3\%$. The delay model rests almost entirely on one feature, no latent dimension is individually interpretable, and 22 chronic line-directions carry half of all delayed trips.
#05Oct 26, 2026
cs.CL
Learning When to Commit: Schedule Distillation for Fast Diffusion Language Models
Aidar Myrzakhan, Zhiqiang Shen
Diffusion large language models (dLLMs) can generate multiple tokens in parallel, but often require many decoding rounds to complete a response. Efficient generation depends not only on which tokens are easy to predict, but also on how committing to them helps predict the remaining positions. A highly predictable token may provide little useful context, while revealing a less predictable token can help several other positions become predictable sooner. We propose dSpark, a new schedule distillation-based framework for learning when to commit tokens for faster diffusion language generation. To achieve this, we construct unmasking schedules that prioritize tokens whose reveal makes other positions easier to predict, reducing uncertainty earlier in decoding. We then distill these schedules into the model to enable faster parallel sampling, with a KL-divergence regularizer to help preserve generation quality. We extensively evaluate our approach across mathematical reasoning and code generation tasks. On GSM8K, our method achieves remarkable 10.12$\times$ and 5.12$\times$ speedups over the various baselines on LLaDA-8B-Instruct and Dream-7B-Instruct, respectively, while maintaining superior task performance. Our code is available at https://github.com/VILA-Lab/dSpark.