Deploying LLMs as agents requires respecting runtime-specific execution contracts, yet standard training leaves harness semantics implicit in action trajectories. This paper introduces Harness Policy Distillation (HPD), which replays reference trajectories through the target harness to compile control-state documents, uses harness verification to filter teacher actions, and jointly trains a shared model on native action prediction, document-conditioned action prediction, and document reconstruction. On TRAJECT-Bench with the Pi harness, HPD lifts Qwen3-8B task success from 3.09% to 22.98% and substantially outperforms supervised, RL, and standard distillation baselines.

Average-reward reinforcement learning with options models continuing hierarchical tasks, but before this paper there were no finite-time convergence guarantees for option-based Q-learning in that setting. The authors prove that multi-step (J-step) inter- and intra-option Bellman operators contract in the span seminorm, design corrected synchronous multi-step estimators, and derive the first $\widetilde{\mathcal{O}}(1/k)$ mean-square convergence rates for average-reward option Q-learning. The results extend to hierarchical reward machines and are supported by small-scale tabular experiments.

This paper studies policy optimization for long-horizon software-engineering agents that learn from sparse binary test outcomes. The core idea is a training recipe—dubbed SPO++—that combines the Muon optimizer with single-rollout asynchronous updates, truncated importance sampling (TIS), token-averaged loss, and auxiliary length/tool-error penalties. The authors claim this combination achieves higher training success than SPO baselines on repository-level issue resolution for two model families, suggesting Muon is critical for long-horizon RL.

Tabular foundation models learn to predict from in-context examples without task-specific fine-tuning, but dense architectures apply the same computation to every feature regardless of statistical context. OrionMoE-Small attacks this limitation by inserting a dataset-conditioned mixture-of-experts layer along the feature axis of a dual-axis Transformer, allowing the model to route different features to different expert networks based on both local token content and a global dataset register. If the sparse design truly helps, it could offer a principled way to scale tabular models by allocating computation heterogeneously; however, the paper never proves the MoE itself is responsible for the gains.

The paper studies how much context a Transformer must retain to perform in-context temporal-difference (TD) policy evaluation without parameter updates. Existing theoretical constructions store the full trajectory, costing Θ(Nd) memory. The authors prove the intrinsic limit is the Krylov dimension k_L, construct a fixed-parameter attention network that compresses raw transitions to Θ(nk_L) registers, and give tight worst-case approximation error Θ(γ^m max{1,1/(m(1−γ))}) when the context falls below k_L.

Explorer-Zero tackles the data-curation bottleneck for agentic RL by training a learned proposer to explore stateful environments, compose multi-step tasks from demonstrated transitions, and record intermediate environment states as verifiable checkpoints. These checkpoints supply fine-grained rewards $R^{\text{subgoal}}$ that prevent group-normalized advantages from collapsing when the solver fails, while alternating proposer–solver training adapts the curriculum automatically. On ALFWorld, ScienceWorld, and AppWorld, the approach outperforms strong self-play baselines and narrows the gap to human-data RL.

Medical image parsing requires a single model to produce heterogeneous outputs—categorical labels, bounding boxes, counts, numeric estimates, and free-text reports—across disparate imaging modalities. This work studies task-specific adaptation of a shared 4-bit Qwen3.5-9B backbone using lightweight LoRA adapters, where a question-only GLiClass router selects the adapter, image budget, and prompt for each case on the FLARE 2026 Task 3 benchmark. The core finding is that adaptation recipes should vary by task—rank, resolution, and prompts allocated according to demand—rather than applied uniformly.

This paper studies the implicit bias of the Muon optimizer when its idealized spectral-norm steepest-descent update is replaced by the standard practical implementation: finite Newton–Schulz iterations with a positive stabilization constant. The authors prove that, contrary to the spectral max-margin behavior established for idealized Muon, the practical variant instead converges in direction to the Frobenius max-margin separator—effectively recovering the implicit bias of ordinary gradient descent. They further characterize how the iteration count, stabilization, and polynomial choice interact, and provide conditions (logarithmically growing iterations) under which spectral bias can be recovered.

Sequence models usually require costly backpropagation and large datasets, leaving open the question of how far simple random reservoirs can go in few-shot settings. This paper systematically tests Echo State Network variants on the STREAM benchmark of synthetic seq2seq tasks, showing that orthonormal input initialization preserves a quasi-linear memory regime and that a proposed data-dependent gate (G-ESN) together with shared deep readouts pushes the solvable frontier toward structural tasks like adding and bracket matching. By benchmarking these cheap reservoir models against gradient-based SOTA, the study offers a practical map of when random features suffice and when fully trained models remain necessary.

Federated fine-tuning lets LLMs learn from distributed data, but malicious clients can poison the global model and their attack patterns may evolve as new tasks arrive. This paper proposes Safe-FCL, which continually adapts a risk detector using historical protective subspaces and then performs risk-aware, module-level aggregation of LoRA updates. The goal is to keep the global model safe across sequential tasks without forgetting previously learned attack signatures.

Neural networks require vector inputs, yet many problems involve higher-order objects like probability distributions, functions, and graphs that lack compact representations. This paper proposes a unified encoder-decoder framework that embeds families of such objects into fixed-size vectors using only finite sample sets, with a domain-specific decoder providing the training signal. The core idea is to replace per-object fitting with amortized learning: a permutation-invariant encoder compresses samples into an embedding that conditions a decoder, enabling generalization to unseen family members. The authors instantiate this on distributions, functions, and graphs, and claim it is the first framework general enough to work across object types.

On-policy self-distillation gives reasoning models dense token-level supervision via a teacher that sees reference solutions, but this signal can drift toward style and ignores whether the student can finish correctly on its own. VASPO repurposes the same privileged reference inside PPO to train both a prefix-value critic and a bounded self-teacher that reweights selected positive advantages. The paper reports average gains of roughly 2.6 percentage points over PPO on math benchmarks and leading accuracy on Knights and Knaves.

Probes on hidden activations are a promising tool for detecting deception and harmful intent, but they assume models cannot volitionally alter their own representations. Baldelli et al. ask whether open-weight LLMs can control their activations through in-context neurofeedback, devising a cross-model writer-subject protocol to disentangle genuine self-access from in-context pattern matching. Across 42 semantically opaque readouts and six safety monitors, they find that free-generation effects are largely explained by sentence-selection competence, while fixed-sentence control is tiny and monitor evasion is essentially absent. The results rigorously bound the current threat but flag a capability that may grow with model scale.

Model selection for time-series anomaly detection remains a critical open problem because no single detector dominates across heterogeneous datasets. This paper proposes RAMSAD, a retrieval-based framework that replaces trainable meta-learners with non-parametric similarity search over a knowledge base of past series and their detector-performance profiles. By matching new queries to previously seen instances via time-series foundation model embeddings—optionally augmented with multimodal semantic alignment—the authors argue they can achieve stronger in-distribution accuracy and out-of-distribution robustness at a fraction of the computational cost.

Agent skills let frozen models adapt through editable external instructions, but optimizing them faces a tension: batching many changes obscures which edit mattered, while judging each edit only by its immediate score risks discarding useful stepping stones. AtomSearch proposes contrastive atomic editing—one primitive rule edit tied to execution evidence—and Skill-State Tree Search (STS), which tests short edit compositions and keeps a composed state only if it beats its primitive child. The hope is to find compact, inspectable skills without greedy myopia.

Training large language models is memory-intensive, and subspace projection methods like GaLore can diverge when stochastic noise dominates the selected low-rank gradient space. This paper proposes SMRC and PMRC, which keep optimizer states in a low-dimensional subspace while adding the orthogonal residual gradient back into the parameter update. The authors prove $O(T^{-1/4})$ convergence rates for momentum SGD and for Muon under a mixed stationarity measure, and demonstrate empirical gains on LLM pretraining and fine-tuning up to $1.3$B parameters.

Population dynamics are often observed only as unpaired snapshots, leaving the evolution between measurements fundamentally underdetermined. This paper introduces Neural Wasserstein Splines (NWSP), a method that learns a continuous curve of distributions by penalizing the intrinsic Wasserstein acceleration rather than the acceleration of individual particles, which is representation-dependent. It couples a neural spline generator with a finite-dictionary projection of velocity and its time derivative, achieving competitive results on synthetic and single-cell benchmarks.

Hindsight-Privileged RLVR tackles the inefficiency of group-based RLVR methods like GRPO, which require multiple rollouts per prompt and lose their learning signal when all sampled outcomes are identical. The core idea is to replace group comparisons with a single rollout per prompt, using the trajectory's own realized future as training-only privilege: a frozen snapshot of the generating policy scores each token under both causal and hindsight views, and normalized Jensen--Shannon divergence between these full-vocabulary distributions weights a historically centered reward. This yields bounded token-level credit without a learned critic or reference solution, training $1.2$--$2.3\times$ faster than GRPO on Qwen3-8B but with task-dependent accuracy tradeoffs.

Momentum is a pillar of modern optimization, yet its benefits in reinforcement learning with verifiable rewards (RLVR) remain unclear. This paper studies a KL-regularized bandit formulation of RLVR and proves that, with exact policy gradients, larger EMA coefficients strictly accelerate convergence within a safe overshoot-free range. Its central insight is that stochastic sampling obscures this gain unless the coefficient scales with the learning rate so that $\eta/(1-\beta)$ stays constant, a principle validated in bandits and in LLM post-training on Qwen models.

Semi-gradient TD with nonlinear function approximation is notoriously unstable because the bootstrap target moves as the parameters evolve. This paper studies the problem in the mean-field regime for two-layer neural networks and identifies a stability regime where entropy-driven relaxation dominates the sensitivity of the moving target. Under the resulting condition, the authors prove exponential last-iterate convergence and quantify the regularization bias, then extend the same mechanism to soft Q-learning.