Model-based reinforcement learning can leverage simulators to avoid risky target interactions, but dynamics shifts create a dilemma: should one correct the simulator or fit the target directly from limited data? This paper introduces MC-WM, which trains both a residual-corrected simulator and a direct target world model, then selects between them using a held-out calibration-risk score. The core value is a disciplined experimental protocol—including exact routing counterfactuals and careful access accounting—that tests whether calibrated model-family selection outperforms fixed strategies.

YouRA tackles a structural reliability problem in end-to-end research agents: manuscript claims drift away from executed experiments because hypotheses, failure histories, and evidence pointers are buried in transient dialogue context rather than kept as persistent, verifiable state. The architecture introduces a YAML-based Verification State Architecture (VSA), an Independent Controller that separates control from execution (using GPT-5.2), and Stateful Reflection with three-level failure routing (bounded repair, redesign, or reset). Evaluation on MLR-Bench's 10-task end-to-end subset reports the highest Overall scores across matched backbones, complemented by hallucination and data-provenance diagnostics, though non-significant differences on the latter versus the strongest baseline temper the reliability claims.

Tabular foundation models like TabPFN achieve strong average accuracy but crash on minority groups when subpopulation proportions shift between training and deployment. DR-TFM addresses this by training only a tiny query-scaling network—updating as little as 0.016% of TabPFN-3’s parameters—using distributionally robust objectives on *estimated* groups, with no true group labels needed. Across tabular benchmarks and multiple TFMs, it delivers large worst-group accuracy gains while staying computationally cheap.

Machine learning models often fail under subpopulation shift when spurious correlations mislead them into relying on shortcut features. This paper introduces POTER, a method that derives per-sample importance weights from optimal-transport (OT) dual potentials between the training distribution and a small validation-based reference distribution. By avoiding loss-based proxy selection entirely, POTER aims to remain robust under label noise while requiring only a single standard ERM training stage. The authors report state-of-the-art worst-group accuracy on several benchmarks and identify a previously overlooked failure mode: subgroup-concentrated label noise.

This paper asks whether Large Language Models can serve as stable, implicit world priors for simulation and decision support when real-world data are unavailable. The authors introduce a framework to evaluate full LLM-generated probability distributions over discrete choice sets—airports, restaurants, and consumer products—across nine open-weight models, three families, and repeated sampling with controlled prompt perturbations. Their central finding is that model scale and family introduce far larger distributional shifts than prompt wording, with modes stabilizing rapidly but encoding structurally different priors across models. This directly challenges the assumptions underlying recent silicon-sampling work that treats LLMs as interchangeable stand-ins for human respondents.

The paper tackles the waste of intermediate reasoning in multi-sample LLM math inference, where self-consistency discards unfinished arguments and only votes on final answers. ReSolve proposes a training-free selective moderation loop: after sampling four candidates, a controller invokes the same 4B model to read existing derivations when answers disagree or are missing, then generates a repaired or completed solution within a bounded budget. The idea matters because it reframes already-generated reasoning as reusable evidence, potentially replacing independent resampling with cheaper evidence-conditioned refinement.

Spacecraft rendezvous and proximity operations currently require specialized engineers to translate high-level mission intent into dynamically feasible trajectories, creating a scalability bottleneck for distributed space systems. This paper introduces OrbitTAMP, a hierarchical task-and-motion planning framework that leverages pretrained LLMs to map natural-language commands into a partial mission specification, then systematically completes unspecified behavior, timing, and waypoint decisions via a domain-specific behavior graph and planning modules. Final trajectory optimization ensures dynamical feasibility and safety, aiming to provide an auditable, scalable language interface for spacecraft operations.

Multi-agent workflows can solve complex tasks that stump a single LLM, but current methods typically train only the workflow generator while leaving execution agents frozen despite their shared influence on outcomes. This paper proposes FloWright, a paradigm that treats the workflow itself as a shared harness to optimize any role—generator, executor, or skill—using a hierarchical, structure-aware reward derived from execution traces, without extra models or labels. It also introduces DataWright, a rule-based data hardening procedure that converts single-agent datasets into workflow-level benchmarks. On 44 evaluation arms across documents, slides, charts, code, math, and finance, the authors show that co-evolving multiple roles yields larger gains than evolving one alone.

Concurrent actions in LLM agent environments can be individually valid yet jointly infeasible. This paper audits settlement semantics through a typed snapshot–settlement contract, testing five arbitration policies for order sensitivity, useful progress, and replay consistency. The core finding is that joint order invariance does not guarantee useful execution: in a six-agent doorway task, priority arbitration raises completion from 31.25% to over 90%, yet still misses the feasibility oracle optimum. The work matters because settlement mechanics directly shape measured agent capabilities but are often left implicit in multi-agent frameworks.

APIVIS addresses the vanishing-learning-signal problem in RLVR mathematical reasoning by mixing direct and Gumbel-searched completions within each rollout group. The core idea is to treat chunk-level search as a policy-improvement operator with a provable monotonicity guarantee, while a selective InfoSFT auxiliary loss preserves training signal when all group returns are identical. Empirical gains on Qwen3-1.7B and Qwen3-4B are consistent but rest on single-seed runs and roughly 3× the compute of direct RLVR.

This paper asks whether the best policy from an RL training run is necessarily one of its visited checkpoints. It introduces policy-space scaling—the idea that completed RL histories contain stronger, constructible policies—and proposes SURGE, which fuses two checkpoints via delta-first spectral decomposition without additional training or per-query inference cost. Across two 1.5B math trajectories and one 7B code trajectory, SURGE exceeds the best native measured checkpoint while using fewer reasoning tokens.

Knowledge distillation risks propagating the reversal curse when a teacher can recall a relation forward but not backward. This paper proposes directional label distillation: frozen teachers score candidate answers in the direction they know, and the highest-scoring candidate becomes the student's training target. On parent-child facts from Wikidata, this yields 13-15 point gains in student open-ended accuracy and shows that verification can provide accurate supervision where generation fails.

Jev-IDS explores whether a System One Model (SOM)—the typed, probabilistic decision model Jev—can serve as a lightweight few-shot classifier for network intrusion detection. Instead of generating free-form text, Jev answers bounded noul and choice questions that return structured attack probabilities and traffic categories, promising lower latency and cost than autoregressive LLMs under label scarcity. This matters because operational NIDS deployments are constrained by per-flow inference overhead, monetary cost, and false-alarm rates, and conventional supervised detectors fail when labeled attack examples are scarce or entirely unseen.

The paper asks when a nominally weaker LLM can reliably audit patches produced by a stronger coding agent. Its central claim is that groundability—the property that evidence exposes an independently checkable claim about missing behavior—matters more than reviewer scale. On a benchmark of 411 real execution-labeled traces and 101 controlled cases, the authors show that weak reviewers given official execution evidence can achieve near-perfect catch and near-zero over-rejection, while a deployable cascade of static checks and generated tests recovers only part of that upper bound. The finding that parameter count is not a monotonic predictor of review quality reframes the oversight problem from building bigger judges to building better checks.

Post-training quantization (PTQ) for LLMs is typically performed sequentially, block by block, by minimizing a reconstruction loss. The authors show that the standard MSE loss creates an Optimization Imbalance: because gradient magnitude scales with the square root of loss, early stages receive much weaker updates than later ones. Replacing MSE with RMSE implicitly normalizes gradients across stages, improving quantization accuracy—especially at aggressive bit widths like W2A16 and W4A4—with no extra parameters or meaningful compute overhead.

JoinGR attacks a real bottleneck in Text-to-SQL pipelines: dense retrievers rank tables independently and miss ones only identifiable through multi-hop join relationships. The method builds a column-level join graph with typed intra-table and foreign-key edges, selects semantic anchor tables via dense similarity, then propagates query-conditioned relevance deposits along edges using a lightweight learned scorer. Final table scores aggregate these deposits, making schema structure a first-class retrieval signal rather than an afterthought, which matters most for enterprise settings with cryptic table names and implicit relational dependencies.

Vision-language-action policies often fail on partially observable, long-horizon manipulation because the current image alone does not determine the correct action. This paper reformulates robotic memory as an optimisation problem: the optimal memory should maximise the conditional mutual information $I(a_t; m_t \mid o_t)$ between action and memory given the current observation. To realise this, the authors propose Divide-and-Remember (D&R), a recursive lightweight selector that retains the top-$K$ most action-relevant tokens from an unbounded history while staying in the pretrained VLA token space. On the RoboMME benchmark and in real-robot experiments, D&R achieves strong gains with budgets as small as 64 tokens, offering both theoretical clarity and empirical progress on memory-augmented visuomotor control.

InFlowOp attacks the costly cycle of building and fixing multi-agent LLM workflows by pricing every decision in a single label-free cost. The framework first constructs a workflow at the right granularity via bidirectional cost-driven coalescing, then detects and repairs faults locally during execution without reference answers. The authors also introduce Braid, a benchmark designed to demand multi-agent coordination, and report accuracy gains up to +11.97% over a single agent across eight domains and six backbones.

Structured argumentation frameworks such as ASPIC- compute whether claims are accepted under grounded semantics by chaining strict rules ($p_1,\ldots,p_n \rightarrow c$) and defeasible rules ($p_1,\ldots,p_n \Rightarrow c$), but interacting with them has traditionally required writing symbolic literals and move primitives. ABDA-NL solves this usability barrier by wrapping the ABDA engine in a natural-language web interface that renders arguments, attacks, and discussion-game explanations in plain English, and by using a large language model strictly as a translator and research assistant rather than a reasoner. Because the deterministic engine alone constructs arguments and assigns acceptance labels, users can explore and extend formal scenarios without losing logical guarantees.

Current multi-agent AI scientists rely on fixed, design-time workflows, yet human scientists adapt their division of labor dynamically. This paper proposes Runtime Agent Coordination (RAC), which lets agents select the next capability at runtime based on research state, scoped work contracts, and artifact-grounded verification. Evaluated across three existing hosts on ResearchClawBench, the key finding is that runtime selection alone improves mean scores, but adding contracts and verification often hurts performance under tight budgets. The work is conceptually sound but empirically thin, resting on a single seed and tiny task samples.