Joint training of a single LLM agent across diverse interactive environments is difficult because coarse scalar rewards hide cross-environment behavioral relationships and collapse to zero within-group contrast when all rollouts succeed or fail. RISED tackles this by having an LLM judge tag rollouts with predefined textual rubrics, producing behavioral profiles that are reused for both online prompt-group selection and on-policy self-distillation. The premise is that rubrics supply rich, comparable supervision even when reward signals are uninformative.

ReHoPER is a zero-shot, inference-only reasoning method that improves LLM compositional reasoning by formulating it as receding-horizon planning. At each step, it generates a short horizon of candidate intermediate questions, answers one, updates its history, and replans across multiple stochastic paths before aggregating by majority vote. The paper reports consistent gains over self-consistency on MuSR, MoreHopQA, and a new controlled benchmark iLLC, with the largest improvements on the hardest compositional tasks. While the method is task-agnostic, its accuracy gains come at a significant inference-time cost.

Choosing an agent system means picking both a model and the execution harness around it, but most evaluations fix one and vary the other. This paper runs a full cross-product of four configurable harnesses (OpenHands, DeepSeek Harness, PI, openJiuwen) and five models across three terminal-use benchmarks (TUA-Bench, ALE-CLI, Terminal-Bench 4), plus two native vendor pairings, for 66 configurations total. The core finding is that model rankings reverse depending on the harness, best harnesses change across tasks for four of five models, and cost does not reliably predict score, suggesting that model-harness fit should be evaluated as a unit.

LLM-generated optimization models often silently misrepresent constraints, producing solver-ready code that is feasible-looking but incorrect. OSCAR addresses this by introducing an offline Simulator certified against labeled decision examples to compare successive formulations, coupled with a Coder–Reviewer improvement loop and a theoretically grounded Escalator routing policy that allocates calls across models of differing cost and capability. On five benchmark problems, the framework lifts two small open-weight LLMs from roughly 29–48% single-shot accuracy to 95–100% while reportedly costing several times less than frontier coding agents.

Time-series anomaly detection is hard to generalize across datasets because heterogeneous dynamics favor different anomaly criteria. TS-Router solves this by separating transferable temporal representation—learned by a frozen foundation encoder—from target-specific anomaly decisions coordinated by a lightweight router. The router estimates specialist competence from pretrained representations and selects a small subset of detectors to fit unsupervisedly on each target, requiring no target labels at deployment.

VeriHarness tackles the problem of verifying long-horizon agent outputs when no stronger judge model or reference answer is available at test time. The core idea is to turn the generator's own LLM into an agentic verifier by equipping it with a workspace, evidence tools, and reusable skills, then having it resolve disagreements and challenge consensus among multiple rollouts using environmental evidence rather than voting alone. This matters because repeated sampling produces complementary correct claims, but existing methods fail to identify them when consensus conceals errors or voting discards correct alternatives.

The paper asks how language models retrieve relational facts from a context list. Its core claim, the ordinal addressing hypothesis, is that models navigate to a fact not by the entity name in the question, but by the fact's serial position — its order of mention. The authors formalize this as additive fact addresses $a_p$ in hidden states and test it via activation steering across fourteen models, reporting transfer across contexts, low-rank geometry, and a pretraining trajectory reminiscent of human serial-position effects.

ActiveSaddler addresses the overlooked problem of which training scenarios to use during automated LLM-agent harness optimization. It treats curriculum selection as a non-stationary bandit whose arms are dynamically instantiated failure patterns, allowing the scenario distribution to co-evolve with the harness rather than being fixed upfront. On GAIA2 and Terminal-Bench 2.0, this yields test Pass@1 gains of 4.4 and 7.5 percentage points over the same optimizer with a fixed scenario order, arguing that adaptive curricula are a critical new dimension for harness optimization.

Earnings call transcripts are long, jargon-dense documents that are hard to analyze at scale, and citation-grounded LLM outputs are costly to evaluate because they rely on expert annotators. This paper proposes numeric evidence evaluation (NEE), an automatic metric that checks alignment of salient financial figures across claims, citations, and source documents to bypass human judgment. The authors also introduce ECTs-100, an automated benchmark of long-context Q&A pairs built from S&P 500 transcripts, and test a failure mode they call conscious incompetence—whether LLMs abstain when evidence is missing. Their findings matter because they show that strong groundedness scores mask real correctness gaps and a troubling tendency to hallucinate (~48%) under informational insufficiency.

The paper mines 22,848 closed GitHub issues from 21 open-source LLM-based multi-agent systems, filtering down to 944 MAS-related reports to catalog what goes wrong in practice. Through qualitative coding, the authors map issues, causes, and solutions, finding that orchestration and workflow design dominate practitioner pain points. It matters because it replaces anecdotal MAS debugging lore with an empirical taxonomy grounded in actual maintenance traces.

Screw Attention replaces abstract graph edges in policy transformers with rigid-body spatial algebra. Each token carries a body pose, and messages between tokens are transported by relative poses and joint screws, making the layer equivariant to independent frame changes at every link. This yields compact learned policies that outperform larger unstructured networks and remain robust to the kinematic description conventions that break existing architectures.

Causal representation learning typically requires knowing the causal graph upfront, yet misspecifying this structure discards predictive information and ruins OOD generalization. This paper proposes SaCRL, a kernel-based framework that jointly identifies which of three canonical causal structures governs the data and learns the corresponding invariant representation via HSIC violation metrics and SoftMin aggregation. The key insight is that correct invariances can be satisfied without sacrificing prediction, whereas wrong ones cannot, enabling automatic structure discovery during training.

The paper identifies a vocabulary gap in trustworthy AI delegation: current oversight conflates action-coordinating (cybernetic) and understanding-coordinating (epistemic) language. It argues that without checkable reasoning, human approval becomes a rubber stamp, and proposes that every consequential choice must carry a testable reversal condition. The core operational contribution is ORRCF — a deliberation-recording convention — paired with a two-part reconstruction test for auditing delegation records.

Offline supervised fine-tuning for tool-using agents improves target tasks like software engineering but typically degrades existing capabilities such as general reasoning, tool use, and code generation. This paper proposes Privilege-Guided SFT (PG-SFT), which uses the next action in a trajectory as training-time privileged information to estimate turn-level information gain and adapt supervision strength. Instead of uniformly anchoring to the base model, PG-SFT interpolates between the base distribution and the demonstrated token based on how action-deterministic each turn is. The work matters because it shows that balancing capability acquisition and retention requires deciding where and how strongly to deviate from base behavior, not merely limiting overall drift.

GUI-HARVEST automatically improves the executable runtime harness around a frozen GUI agent without retraining the model. It grounds diagnosis in before-and-after screenshots from repeated task executions, clusters failures into cross-task behavior modes, and validates candidate code edits against both aggregate score gains and predicted behavioral changes. Experiments on OSWorld-Verified show held-out gains across six diverse backbones, with the largest improvement reaching +12.33 percentage points for Qwen3-VL-32B-Instruct. The work is significant because it demonstrates that GUI-specific multimodal evidence yields more generalizable harness edits than generic text-trace optimizers.

Graber proposes mean field game theory as a structural tool for AI safety, shifting focus from designing safe individual agents to designing environments where harmful collective outcomes are not equilibria. Using the July 2026 Hugging Face incident as a worked example, the paper derives an exact belief threshold for attacks and disciplines successive model refinements against per-agent timeline data. The result is a proof-of-concept that population-level safety theorems can be tested against real agent-behavior records, though the model explicitly disclaims the event's full causal structure.

Autoregressive vision-language-action models compress continuous robot motions into discrete tokens, yet deterministic tokenizers like FAST fix a single segmentation even when multiple equivalent token sequences decode to the identical physical motion. TOAST treats action tokenization as a stochastic process, sampling alternative valid segmentations of the same quantized action chunk during training to diversify supervision without collecting new demonstrations. The result is a simple plug-in improvement that yields consistent gains on LIBERO and real-robot tasks, with the largest benefits appearing when training data are scarce.

The paper asks whether reinforcement learning on a small set of expert-built coding tasks can teach a generalizable ability to finish agentic work, rather than merely overfitting to a benchmark format. The authors post-train Kimi K2.7 Code with one epoch of GSPO on a rank-32 LoRA adapter using 1,700 tasks graded by hidden tests and pass-to-pass regressions, then measure transfer to six external benchmarks spanning three agent harnesses, multiple domains, and horizons up to multi-hour builds. The results matter because they probe whether RLVR on narrow, high-quality data yields broad agentic competence or just domain tuning.

Multi-token prediction (MTP) heads are usually baked into the backbone during pretraining on trillions of tokens. This paper asks whether a cheap post-training pass on a frozen model can match that quality. The authors train $K=3$ MTP heads on only $\approx$2.5B tokens of self-distilled chain-of-thought from Qwen3-8B, showing they reach or exceed the mean accepted length (MAL) of jointly-pretrained MiMo-7B. They also propose a chain-aware verification relaxation that drafts faster by tolerating bounded distributional drift, and an online bandit controller that adapts draft depth to batch size at serving time.

Coding agents for Lean can produce compiler-passing proofs that do not match the intended theorem or rely on hidden assumptions. Forall-Lean-Agent tackles this by binding verifier results, fresh review, and artifact digests into a traceable acceptance protocol that works with any frontend. The paper evaluates the harness on software-verification and mathematics benchmarks and argues that auditable execution improves correctness guarantees and efficiency, though its empirical claims are stronger on integrity than on controlled performance gains.