In briefOff-policy RL with online inference tuning for the CoRL 2026 RoboEval challenge
SummaryThis paper describes the author's solution to the CoRL 2026 RoboEval challenge, which has eight bimanual tasks scored on task stage completion, safety, efficiency, and arm coordination compared against human demonstrations. The method is off-policy reinforcement learning on a π0.5-based policy that also serves as its own value function, with each rollout doubling as a Thompson-sampling trial over inference parameters. The submitted policy reaches 0.990 RDS on held-out seeds and 0.987 on the official seeds.
In briefAligning unpaired model representations using a shared pullback distance on the data manifold
SummaryThe paper introduces Neural Gromov–Monge Alignment (NGMA), a method that aligns two models' representations without paired data or labels. It uses the pullback of the data metric as a distance shared by any two models of the same data-generating process, under a precise manifold hypothesis. The alignment map is identifiable exactly when the data distribution breaks every symmetry of the data manifold, which the authors show holds generically. NGMA outperforms unpaired and label-informed baselines on image and single-cell models.
Stas Syrota, Albert Lopez I Serrano, Johanne Franck et al.
In briefMOMENT-DP: Momentum-Aware Evaluation of Correlated Noise for Private Training
SummaryThe paper introduces MOMENT-DP, a framework for evaluating temporally correlated noise used in differentially private optimization. It models how privacy noise propagates through momentum-based optimizers, accounting for curvature, learning-rate schedules, momentum, and noise covariance. The authors show that changing only the momentum can reverse the ranking of privacy noise mechanisms, indicating that optimizer state is an important factor in evaluating correlated privacy noise.
In briefSemantically aligned KV cache compression for efficient agentic video generation
SummaryThis paper introduces SemanticKV, a training-free method that compresses the key-value cache in agentic autoregressive video generation. It uses semantic constraints from the planner, identified through text cross-attention, to score cache entries and keep the high-scoring ones under a memory budget. On Causal Forcing at a 0.7 compression ratio, it reaches a VBench Total Score of 84.10, above Full KV's 83.99, with a 1.22x speedup and KV-cache memory reduced from 6.04 GB to 1.81 GB.
In briefHow much train-inference mismatch on-policy distillation can handle
SummaryThis paper studies how much mismatch between the training and rollout inference engines on-policy distillation (OPD) can tolerate. The authors measure all mismatch sources on one scale, testing quantization across three student-teacher pairs and weight staleness and injected noise on one pair. OPD tolerates substantial mismatch without meaningful quality loss but degrades beyond a threshold. They also propose a lightweight admission check against a calibrated budget, along with model-specific FP8 recipes and asynchronous rollouts that improve efficiency.
In briefLearning How to Rank Constraint Propagators for Faster Solving
SummaryThis paper proposes learning instance-dependent priorities for propagators, the components that narrow down variable domains in constraint programming solvers, instead of using fixed manual heuristics. A self-attention model scores propagators and is trained on pairwise preferences derived from solver runtimes under randomized priority assignments. On 698 difficult XCSP3 instances, compared with Choco's default priorities, the learned priorities raise solved instances from 634 to 654 and cut median runtime by 27.93%.
In briefSWHForest: A structural layer for filtering large code corpora without redistributing source code
SummaryThe paper presents SWHForest, an open-source framework and public dataset that adds a structural filtering layer to code corpora published only as file pointers. It parses each file once into an abstract syntax tree, discards the source text, and keeps only structural summaries so that matching files can be selected before downloading. On the ten largest languages of Stack-Edu (52.1M unique files, 341M definitions), queries answer in milliseconds and cut downloads by 46-72%.
Federico Pennino, Andrea Gurioli, Maurizio Gabbrielli et al.
In briefBenchmarking general AI agents for understanding surgical scenes
SummaryThis paper evaluates general-purpose multimodal language models on surgical scene understanding, which is hampered by scarce data and inconsistent evaluation methods. It introduces SurgKnowBench, a diagnostic benchmark built from harmonized public surgical datasets, showing models are weak in relational, procedural, and workflow understanding. It also proposes SurgAgentGym, an environment where agents use bounded knowledge search and visual grounding at inference time without surgical training. Results indicate agentic inference helps mainly perception-heavy tasks, while fine-grained localization and expert-style assessment remain bottlenecks.
