Tool use, function calling, orchestration and the protocols that let models act rather than only answer.
arXiv:2610.10513v1 Announce Type: cross Abstract: Language-model agents are increasingly asked to carry out open-ended scientific research, yet their results are usually graded against a known answer...
arXiv:2610.09285v1 Announce Type: new Abstract: Many applications of physical AI run within finite or closed physical worlds with a limited set of physical laws governing object behavior. Examples in...
The study compares supervised fine-tuning (SFT) and reinforcement learning (RL) as teacher training methods for on-policy distillation (OPD) in a single-teacher setting. Using Qwen3.5-9B models across Agentic, Reasoning, and Perception domains, RL-trained teachers consistently produce stronger students and better recovery of performance gains than SFT-trained teachers. The authors attribute this advantage to RL teachers’ smaller parameter shifts from the shared initialization, which facilitates learning through OPD.
The paper introduces Task-Progress Distillation (TPD), an offline method that pairs each action in teacher demonstrations with a short label indicating the current task stage. A small 1.7B student model trained with 404 demonstrations using TPD achieves 72.4% mean unseen task success on ALFWorld, matching action-only supervision and outperforming a reasoning-trained student. At lower demonstration budgets, explicit stage labels significantly boost performance, raising success from 48.0% to 67.7% with 200 demos, and both approaches converge to 76.9% with 808 demos.
arXiv:2610. 10527v1 Announce Type: cross Abstract: Heavy-tailed noise has been widely observed in modern machine learning, motivating the use of methods like gradient clipping and normalization.
ActiveLang is an autonomous system that builds language‑annotated 3D maps by actively exploring environments. It uses a compact dual‑Gaussian representation to jointly reconstruct geometry, appearance, and open‑vocabulary semantics while adapting language features online. The planner selects informative viewpoints, reducing the number of observations and computational cost, and experiments on Replica and ScanNet++ show significant gains in 2D and 3D open‑vocabulary segmentation compared to existing baselines.
The paper investigates how randomized exploration in bandit learning can be compromised when an adversarial agent gains access to another agent’s exploratory signals. In a two‑player strongly monotone setting, the deceptive agent couples its actions with the leaked information to steer the learning dynamics toward a new steady state, termed the deceptive Nash equilibrium (DNE). The authors prove that the deceptive bandit learning dynamics converge to an arbitrarily small neighborhood of the DNE while maintaining optimal convergence rates, even under relaxed smoothness assumptions, and illustrate the impact on costs in a resource‑allocation game.
The paper presents Context-Aware Attention-based Gaussian Mixture Models (CAA‑GMM) for predicting vehicular trajectories in complex, uncertain environments. CAA‑GMM models future motion as a probabilistic mixture conditioned on scene context and agent dynamics, using a lightweight attention mechanism to encode inter‑agent interactions and contextual salience. Experiments on nuScenes and Argoverse 2 show competitive or superior accuracy to raster‑based baselines while keeping computational complexity low, and ablation studies confirm the attention module’s importance for robust contextual reasoning.
The paper introduces temporally interpretable differentiable decision trees (DDTs) that incorporate action chunking to better align tree-based decision making with multi‑timestep planning. Two new policy‑gradient algorithms and an information‑theoretic tree‑restructuring method are proposed to enable efficient training of these trees. Experiments on four simulation environments show that warm‑starting action‑chunked DDTs from a distilled policy yields trees that match neural network performance in three domains while using up to 80% fewer parameters.
The paper introduces Continual Graph Multi-Agent Reinforcement Learning (CGMARL), a framework that maps sequences of tasks into attributed graphs to capture task-specific structures such as network topologies or formation arrangements. It presents the Graph-based Formation (GRAFO) benchmark to illustrate forgetting in graph-based CMARL settings and proposes the Frozen Graph Encoder (FROG) to preserve past structural information using a frozen graph backbone. Experiments on GRAFO demonstrate that combining FROG with existing continual learning methods significantly improves performance across multiple CGMARL scenarios.
SkillRL is a framework that enhances large language model agents by automatically discovering and evolving skills from raw experience. It builds a hierarchical skill library called SkillBank, uses an adaptive retrieval strategy for heuristics, and allows the skill library to co‑evolve with the agent’s policy during reinforcement learning. These techniques reduce token usage and improve reasoning, achieving state‑of‑the‑art results on ALFWorld, WebShop, and seven search‑augmented tasks, outperforming baselines by 15.3% and remaining robust as task complexity grows.
SPW-Nav is a streaming panoramic world model that generates one minute of 2K 360‑degree video in real time from a single panorama, guided by language instructions. It interprets each instruction as camera motion, applying spherical rotation decoupling, pose‑aligned conditioning, and a multi‑term memory to maintain continuity as instructions change. The authors also release SPW-NavSet, a dataset of panoramic videos with camera trajectories and verified instructions, and demonstrate that SPW-Nav surpasses prior panoramic generators in camera‑following accuracy and video quality while enabling on‑the‑fly instruction switching.
trACT (temporal revelation Airborne Camera Trap) is a lightweight, real‑time aerial robotics framework for autonomous consumer drones. It uses Temporal Max Pooling to convert subtle ground motion into robust signals, coupled with self‑supervised motion anomaly detection to separate target movement from environmental noise. The system also incorporates motion prediction, gimbal‑stabilized optical zoom verification, and balanced multi‑target verification, enabling precise target confirmation in challenging forested and surveillance environments.
The paper introduces Ctrl-CWM, a multi‑agent controllable crowd world model that merges crowd generation with run‑time control. It uses an encoder to learn human motion dynamics, an actor to propose pedestrian displacements, a critic to evaluate imagined trajectories, and a planner that selects actions based on critic scores and user‑defined costs. Experiments show that Ctrl‑CWM outperforms state‑of‑the‑art methods in realism and collision metrics while allowing new control objectives to be added without retraining.
GraphOPD introduces a graph‑augmented on‑policy distillation method for large language model agents, addressing the shortcomings of existing divergence‑based guidance when rewards are sparse. By constructing a dependency graph from the environment’s state‑change record and applying a random‑walk stationary distribution, the method identifies and prioritizes the most impactful steps in a trajectory, fusing this structural credit with teacher‑student divergence signals. Experiments on ALFWorld, WebShop, and SearchQA demonstrate that GraphOPD consistently outperforms eleven baselines, improving the strongest baseline by up to +5.8 percentage points and showing strong causal alignment through an executed‑replay audit.
SPIN (Shadow Predictive Indexer) is a new method for sparse attention that reduces the overhead of scoring the entire key-value (KV) cache during decoding. By using lightweight, history-based predictions to identify important KV blocks, SPIN avoids full cache scoring at each step and treats KV blocks and speculative decoding as core design elements. Experiments on long-context and agentic benchmarks show 30‑40% sparsity with maintained task quality, and in vLLM serving it boosts throughput by up to 14.9% while cutting median inter‑token latency by up to 13.2%.
TAP (Trajectory‑Anchored Pruning) is a structural pruning framework designed for reinforcement‑learning‑trained agents that addresses the shortcomings of existing pruning methods on long‑horizon tasks. It couples iterative channel pruning with on‑policy recovery, anchoring student interactions to teacher trajectories and re‑scoring channels using gradients from the recovery objective. Experiments show that removing 60% of FFN channels preserves 99.2% and 88.0% of task success rates on ALFWorld and WebShop while cutting GPU time per successful task by about 22% and 17%.
VideoEvolve is a self‑evolving framework for long‑video understanding that jointly updates both memory and retrieval mechanisms. It starts with a coarse low‑frame‑rate overview and uses a Memory Evolver to selectively augment memory while a Retrieval Evolver adapts how information is fetched from that memory. The two components are alternately refined via Agentic Reinforcement Learning, guided by Bottleneck‑Aware and Capability‑Aware Evolution Feedback to focus on the most limiting aspect and to prevent over‑specialization, respectively. Experiments on several long‑video benchmarks show that this approach improves performance over static systems.
The paper introduces a hybrid quantum‑classical framework for detecting affinity and romance‑investment fraud by modeling cognitive biases in manipulative conversations. Dedicated qubits encode key biases, a frame qubit captures temporal order, and a narrative qubit aggregates co‑occurrence via a trainable entanglement layer. The circuit parameters are jointly trained with a classical reinforcement‑learning agent that decides, turn by turn, whether to flag the conversation as an optimal stopping problem, and the model is evaluated on synthetic conversations including hard negatives and legitimate but urgent or pushy sales scenarios.
The paper introduces Interaction-Constrained Drive Policy (ICDP), an offline reinforcement learning framework for autonomous driving that explicitly controls interaction-level distribution shift. ICDP decomposes joint-support degradation into ego-support and interaction-support components, estimating the latter via contrastive density-ratio estimation without requiring joint-density modeling or simulator rollouts. Experiments on nuPlan, Interplan, and real-world truck data demonstrate that ICDP reduces selection of high-value but interaction-unsupported trajectories and improves performance in interaction-critical driving scenarios.