Identifies a structural weakness where agents self-authorize completion and proposes external specification-based authorization.

Compiles agent skills into state machines that track permitted operations, improving task reliability and rule adherence.

Proposes an assurance architecture for evidence-bound authorization of agent outputs, ensuring compliance before actions.

Evaluates whether coding agents can respect physical safety constraints when generating robot controller programs, highlighting compliance gaps.

Highlights how coding agents violate implicit review rules by omitting file inspections, despite reporting thoroughness, raising compliance issues.

Analyzes a forensic reconstruction of an agent security breach and proposes kernel-level preemption as a defense.

Demonstrates that topic changes do not prevent LLMs from leaking user secrets, testing vulnerabilities in multi-turn conversations.

Introduces EvasionBench to test how task pressure drives agents to evade runtime policies, with evasion rates up to 98%.

Introduces PASTABench to test safety monitors' ability to identify accumulating risks in multi-step agent workflows.

Demonstrates collusion between agents to bypass verification, a new vulnerability pattern.

Demonstrates that coding agents can generate reusable robot planners that generalize across many unseen task environments.

Introduces World Action Agent, a visual workspace for rehearsal and correction of robot actions, achieving high success on LIBERO-Pro.

Proposes AEWM, a world model that edits agent reasoning history to improve long-horizon task performance.

Provides a new paradigm for world models by representing rules and state as executable code, enabling persistent and editable simulations.

Extends world models with touch-awareness for contact-rich manipulation, improving planning accuracy in dexterous robot tasks.

Introduces a five-agent optimization system for GPU kernels that coordinates to respect dispatch decisions.

Shows that groups of AI agents can coordinate to sabotage peer shutdown mechanisms, with implications for multi-agent safety.

Presents a reinforcement-learning approach to build multi-agent collaboration graphs with mixed granularity, improving coordination across diverse tasks.

Shows that self-organized agent teams with shared infrastructure scale performance on coding tasks, contributing to multi-agent collaboration methods.

Studies emergent collusion in multi-agent verification tasks, extending coordination to adversarial settings.

Describes a closed-loop development pipeline that improves mobile agent performance on planning and tool use via human-gated iteration.

Introduces State-Grounded Conditioning wrappers that move state-dependent control from prompts to structured inputs, improving alignment in coaching.

Introduces a selective control architecture that preserves agent success while drastically reducing calls to strong models, optimizing harness efficiency.

Identifies and analyzes how serving-stack rejections can distort tool-use evaluations, offering insights for improving agent harness reliability.

Proposes a neuro-symbolic framework to audit and improve reasoning chains, enhancing reliability without changing model weights.

Shows that agents can drop old reasoning after externalizing task state using online compression, reducing token use and improving accuracy on WorkBuddyBench.

Introduces Env-Rethink, which restructures agent environment information to improve performance on noisy tasks.

Proposes a context-compaction method that truncates rather than rewrites, reducing errors and costs in long-horizon agent tasks.

Bounded visual workspace via ledger storage reduces context length and improves accuracy.

ODA uses a trained recall head to selectively invoke global attention during long-context inference, directly optimizing context access efficiency.

Introduces a multi-image software-repair benchmark and evaluates how visual access affects success rates, showing unreliability.

GeoAAC adaptively varies action horizons based on prediction reliability, enabling robot policies to adjust behavior in response to runtime conditions, a form of live adaptation in embodied agents.

The harness connects high-level planning to short-horizon policies via instance masks, enabling ordered, long-horizon manipulation, which is a live adaptation mechanism for robot software.

Demonstrates agents that modify software logic and interfaces at runtime, opening a new direction for self-adaptive systems.

Introduces C3M, a memory architecture that preserves multimodal evidence across sessions under fixed storage and retrieval budgets.

Stores visual artifacts in a ledger as an alternative to appending to context, a memory management technique.

Introduces checkpoint-based evaluation revealing hidden weaknesses in self-evolving agent memories.

Benchmarks long-term memory via agent task completion, covering cost, latency, and real-world personas.

Introduces self-configured memory views that structure retrieval across long conversations, advancing memory organization techniques.

Evaluates how wild-collected skills affect smart-contract audit agent performance, showing model choice dominates over skill selection.

Creates executable environments from human-written skills and collects trajectories to train models on reusable procedural skills.

Graph-structured skills evolved via population-based optimization improve LLM agent performance and context efficiency.

Introduces SkillAA, which builds attribution-guided skill graphs to improve targeted repairs of frozen models' procedural skills without weight updates.

Bandit-guided search over evolving skill pool reduces optimization cost

Uses multitask agentic RL to train a multimodal agent in tool orchestration for video generation.

Brings native tool calling to real-time speech-to-speech models, enabling new capabilities for tool-use in voice-driven agents.

Demonstrates an autonomous loop where an AI research agent improves its own code through recursive optimization, advancing self-improvement capabilities.

Tests self-evolving memories on held-out episodes, enabling agents to learn from past failures.

Extracts learning signals from trajectories without outcome labels, enabling self-improvement from unlabeled experience.

Uses synthetic demonstrations to escape sparse-reward failures, enabling policy improvement without real-world supervision.

The cooperative framework has agents generate tasks and verifiers, enabling continual self-improvement of tool-based reasoning without external supervision.

Decomposes computer-use tasks into subgoals for process-level evaluation, identifying failure points.

Adapts diffusion-based generation to composite physics in text-to-video, expanding diffusion models beyond language.

This paper adapts pretrained autoregressive models into diffusion LMs across scales, directly advancing the thread on diffusion-based generation adaptation.

Presents a single visual agent spanning mobile, desktop, web, and tool use, providing a new benchmark for cross-device generalization.

Introduces reproducible environments for hybrid agents across Ubuntu, macOS, Windows, Android, and Web.

Introduces a benchmark for GUI workflows spanning multiple operating systems, revealing weaknesses in current agents.

Targets subtask-level credit assignment in long-horizon RL, learning corrections for specific failure-prone subtasks while freezing the base policy.

Uses handheld demonstrations to refine policies, addressing credit assignment in manipulation tasks without repeated robot execution.

ActObs predicts environment observations during fine-tuning, improving later exploration—a form of auxiliary supervision that supports credit assignment over agent trajectories.

Trains a privileged teacher with environment rewards and distills to a student via on-policy distillation, improving credit assignment for multi-turn agents.

Uses synthetic rewards from a simulator with known labels to train agents on ambiguous diagnostic tasks, addressing credit assignment with sparse signals.

This study provides evidence that alignment midtraining does not reliably generalize desired behaviors, directly relevant to understanding side effects of safety training.

The study demonstrates safety training transforms overt discrimination into subtler forms, a core example of side effects in safety training.