#01Oct 23, 2026
cs.RO
When Human Corrections Travel: Safety in Federated VLA Adaptation
Vigneshwaran Subbaraju, Obaidullah Zaland, Monowar Bhuyan
Vision-language-action (VLA) policies can adapt to local human intervention, while federated learning can combine distributed robot experience without pooling raw records. We examine their prospective intersection: sharing adaptations derived from diverse human corrections across tasks, environments, people, and robot embodiments. Corrections differ in modality, safety semantics(urgency, behavioral scope, etc.), and privacy sensitivity. A shared update that improves mean task success may suppress a rare safety intervention or produce an unsafe response on another robot. We argue for three requirements: preserve the context and uncertainty of multimodal corrections when deciding what to share; assess candidate updates in each receiving robot's closed loop before deployment; and evaluate leakage alongside retention of safety-critical evidence. Immediate, legible defensive behavior remains necessary during human interaction. We offer counterexamples and a falsifiable evaluation agenda, not experimental results.
#02Oct 23, 2026
cs.LG
Multiplicative vs. Additive State Coupling in Learned Optimizer Schedules
M. Harper Langston, Pierre-David Letourneau, Dalton Jones and 7 more
We learn schedules for a continuous penalty relaxation of MAX-CUT by backpropagating through an unrolled classical solver. A state-blind schedule generalizes to instances substantially larger than those used for training. A state-adaptive bilinear schedule also generalizes out of distribution, outperforming its untrained baseline across 15 seeds ($t=4.1$). By contrast, an additive ablation with identical inputs and output heads is statistically indistinguishable from the bilinear model in distribution but performs worse than its own baseline out of distribution ($t=-11.7$). This separation appears without regularization and persists across a $30$ round weight-decay sweep: the bilinear advantage remains significant at every tested value, whereas the additive model never recovers. The results in this setting isolate how solver state enters the schedule, rather than state adaptivity alone, as the architectural distinction associated with out-of-distribution transfer.
#03Oct 23, 2026
cs.SE
Trusting the Tool: Asymmetric Deference of LLM Vulnerability Auditors to Static-Analysis Evidence
ZhiHong Liu, Shuyao Gao
Pipelines that combine large language models (LLMs) with static analysis let the model consume the analyser's output as evidence, but whether the model distinguishes true from false evidence, and whether its scrutiny depends on what the evidence claims, has not been measured. We built an intraprocedural analyser for C/C++ that produces proof obligations and audited vulnerable and patched functions from PrimeVul under six evidence conditions that differ only in the notes, including fabricated assurances that a missing check exists and fabricated alarms that an added check is missing. The study covers 916 pairs in four disjoint sets, eleven model configurations from four vendors and 16,572 audits; no model is trained or fine-tuned. True evidence barely improved accuracy (+0.02). False evidence acted asymmetrically: relative to true notes about the same operations, false assurances lowered the detection of real vulnerabilities by up to 19 points, whereas false alarms never significantly lowered accuracy and were disputed far more often. An internally registered replication on seven new models confirmed that false assurances move verdicts towards their claim more than false alarms do (mean asymmetry +0.16, 95% CI +0.10 to +0.24). Instructions in the audit prompt did not reliably reduce deference; the best one failed to replicate. Verifying each claim in a separate call and showing the refuted claims raised accuracy under false assurances by 12 points in a pre-specified test (p=0.002) without loss under true evidence; for a second model the gain was smaller and not significant. Pipelines should therefore verify a tool's guarantees before the model relies on them, and studies should measure deference in a direction-aware way.
#04Oct 23, 2026
cs.LG
Reinforcement Learning with Multi-Step Transition Predictions: A Bellman Theory
Chenbei Lu, Zaiwei Chen, Tongxin Li and 2 more
Multi-step predictions of future transition realizations can improve reinforcement learning, but their value remains difficult to quantify. Directly including these predictions in the state can lead to exponential state-space growth, making direct dynamic programming computationally expensive. We study discounted finite MDPs under a block-feedback protocol: at the start of each K-period block, the decision maker observes $K$-step probabilistic transition predictions for a subset of actions and adapts decisions to observed states within the block. We show that the optimal value satisfies a Bayesian Bellman equation on the original state space. This representation reveals that predictive information creates value by recursively changing the local order of maximization and expectation in the Bellman expansion. We formalize this effect through a Bellman--Jensen analysis and quantify the effects of prediction horizon, error, action coverage, and within-block feedback. We propose block-feedback Bayesian offline learning with online adaptation (BF--BOLA), a model-based algorithm for learning near-optimal policies, and establish upper and lower sample-complexity bounds that match up to logarithmic factors in the stated regimes. The results show that predictions can reduce transition learning to uncovered rows and that learning a near-optimal policy can require fewer prediction samples than estimating the Bayesian value function.
#05Oct 23, 2026
cs.LG
Where Rounding Happens Matters: Rounding-Location Alignment Stabilizes LLM Reinforcement Learning
Yijing Ke, Tianle Cai, Ruosong Wang
Reinforcement learning for large language models often uses separate systems for inference and training. Even with identical weights and BF16 tensors, implementation differences can make an intended on-policy update off-policy and destabilize training. Existing remedies align the entire inference and training stacks, change global precision, or modify the learning algorithm. We ask whether aligning selected parts of the execution stack can stabilize BF16 RL while other implementation differences remain. We align training-side rounding locations with inference in three regions---RoPE, SwiGLU, and residual RMSNorm---by retaining selected intermediates in FP32 until their BF16 outputs. Across four Qwen-family models, aligning all three regions prevents the collapse observed under the default BF16 baseline; RoPE-path alignment alone also suffices in the studied runs, despite remaining numerical discrepancies. The interventions keep the inference engine, matrix-multiplication and attention kernels, and learning objective unchanged.