Tool use, function calling, orchestration and the protocols that let models act rather than only answer.
LiDAR-based 3D object detection is a core perception service for autonomous driving, but deployment requires robustness to point-cloud corruption while meeting latency and memory requirements. Existin...
Recursive self-improvement (RSI) promises AI systems, such as self-evolving LLM agents, that become better at producing further improvements. Evaluating this promise requires separating product claims...
Code agents increasingly work on evolving real repositories, and memory systems aim to help them reuse earlier experience. However, such experience is diverse and evolving, so it remains unclear how m...
On-policy learning is central to post-training: training on a model's own generations limits distribution shift and helps preserve existing capabilities. On hard problems, however, on-policy RL provid...
On-policy distillation (OPD) can succeed with surprisingly few training questions, motivating a view of questions as carriers of distributional knowledge transfer. How can we assess whether a newly ge...
Simon Willison reflects on how programming used to be a rewarding career, describing it as a job where developers solved complex logic puzzles for high pay and had full control over strategy and timelines. He notes that coding agents and large language models (LLMs) now handle many of these puzzles, offering a "gift that keeps on giving" for those who enjoy building systems without caring deeply about the underlying problems. However, for programmers who found joy in directly solving problems, this shift represents a significant loss.
While LLM agents can generate visually plausible CAD assemblies, these designs often fail to fulfill their intended physical functions. We study functional CAD design, a setting where agents interpret...
In this paper, we systematically investigate the mechanisms, scaling behavior, and generalization of physical Recursive self-improvement. We introduce PhysRSI-eval to quantify physical RSI through a f...
Long-horizon embodied agents face temporal confusion and incomplete or distorted records, recurring errors from fixed memory rules, and growing storage and retrieval costs. We present BetaMem, a frame...
Referring Expression Comprehension (REC) localizes the image region described by a natural language expression, underpinning applications such as autonomous navigation, human-robot interaction, and me...
CastClaw is an interactive forecasting system that combines an autonomous forecasting loop with a forecasting-oriented harness. Users define and refine forecasting tasks in natural language via command-line or Web interfaces, while an internal large language model agent selects tool actions or requests human feedback, using checked results to guide the forecasting process. The system integrates tool execution, state management, and validation, and has been evaluated on five public electricity‑price datasets, outperforming 16 baselines in point‑estimate MSE and MAE, with case studies on Nord Pool and North China electricity‑load data.
The paper proposes a centralized training with decentralized execution framework for multi‑agent UAV pathfinding in stochastic environments, leveraging multi‑agent reinforcement learning enhanced by model predictive control concepts. Agents communicate only with a centralized planner and use a distance‑weighted mean‑field observation to plan feasible paths, while multi‑step value convergence improves training efficiency. Experiments, including ablation and real‑robot studies, demonstrate the method’s effectiveness and generalization.
RealCLI is a new benchmark for evaluating human‑agent collaboration in command‑line interface (CLI) tasks. It unifies task execution and collaboration quality assessment through a five‑dimensional framework with 22 sub‑dimensions and 51 interaction templates. The benchmark offers two modes—REALCLI‑EXEC with full specifications and REALCLI‑COLLAB with a scripted user providing progressive feedback—and reveals significant capability gaps in current agents, especially in steerability, trust calibration, and safety management.
The paper investigates how large language models decide whether to call external tools before the call is made, a step that has been largely overlooked in tool‑safety research. Using factorial interventions on input requests, the authors find that neutralizing the requested operation significantly reduces harmful admissions, and that a linear direction derived from this contrast predicts admission across most models. They also show that training models to refuse in plain text can lower unsafe admissions, indicating a partial link between written refusal and tool‑admission decisions.
The paper investigates whether an elicitation agent (EA) can improve the quality of Lean specifications generated by large language models (LLMs) through interactive clarification. The EA decomposes a task, identifies unresolved decisions, asks a limited number of clarification questions, and refines the natural‑language specification before an LLM formalizer translates it into Lean. Experiments on 189 Verina Lean 4 tasks show that cross‑model pipelines (e.g., Sonnet 4.6 → GPT‑5.6) can achieve up to a 65.6 % strict pass@1 rate, whereas same‑model pipelines do not consistently improve accuracy.
The paper introduces the concept of Persona Collapse, where large language models (LLMs) produce clustered, non-diverse responses in multi-agent simulations. It defines three diagnostic metrics—Coverage, Uniformity, and Complexity—to quantify this phenomenon across eight LLMs on tasks such as personality, moral reasoning, and self-introduction. Findings show that persona collapse is universal but varies by task, cannot be alleviated by sampling more personas, and that higher per-persona fidelity leads to more polarized populations, while persona-targeted fine-tuning further reduces diversity.
AgentOdyssey is a new evaluation framework that procedurally generates open-ended text games with rich entities, dynamic worlds, and long-horizon tasks to test test-time continual learning agents. The framework measures game progress and provides diagnostic tests on world knowledge acquisition, episodic memory, exploration, action diversity, and model cost. Experiments across six games show that even the best agents fall far short of human performance, and supervised fine-tuning leads to catastrophic forgetting, highlighting the need for better parametric continual learning methods.
Expand-and-Prune is a test‑time framework for candidate‑based end‑to‑end autonomous driving planners that dynamically expands and prunes the set of trajectory hypotheses. It adds new candidates guided by semantic driving intent and temporally propagated predictions, then uses uncertainty in intermediate trajectory scores to prune the expanded set. Across four planners and four benchmarks, the method consistently improves planning performance, achieving state‑of‑the‑art results on LangAuto and significant gains on Bench2Drive, Fail2Drive, and NAVSIMv2, especially in out‑of‑distribution scenarios.
The paper introduces MemCompile, a query‑conditioned evidence compilation framework designed to improve long‑term conversational memory. It uses an Address Plane to locate relevant support via compact descriptors and capacity‑aware cells, then normalizes retrieval hits into stable SourcePacket identities. An Evidence Plane allocates a bounded evidence budget across packets, admits limited additional support, and produces answer‑ready context, achieving 80.52% Judge Accuracy on LoCoMo and improving the quality‑cost Pareto frontier.
The paper introduces RecReg, a safe reinforcement learning framework that enforces non‑Markovian regular safety properties through recovery‑based shielding. It provides theoretical foundations and rigorous safety and convergence guarantees in the tabular setting, and demonstrates empirical effectiveness using the MASA‑Safe‑RL library. The authors also outline an extension of RecReg to deep RL with state‑of‑the‑art model‑based methods.