Core research on what makes a language agent trustworthy at scale — how it remembers, how it verifies what it believes, how it improves itself from experience and external knowledge, and how it runs cheaply enough to matter in production. World models run through all of it: when an agent can simulate the consequences of an action, feedback stops being a guess.
Reliability & Verification
Memory, belief, and detection mechanisms that make agent behaviour reproducible, auditable, and verifiable against evidence.
CounterMem: World-Model-Verified Counterfactual Memory for Language Agents
Under reviewICLR 2027 · arXiv:2609.31874
An RL framework that builds memory out of verified counterfactuals: it generates local alternatives to failed actions, checks their outcomes against executable world models, stores the action contrasts with evidence and reuse conditions, and trains a DQN selector to decide when retrieved memory should steer a new task. Improved ReAct and Reflexion in 24/24 benchmark-agent comparisons (+12.6 pp average) while reducing task-run token usage.
BeliefRAG: Making Adaptive RAG State-Aware Under Evolving Evidence
Under reviewTACL · arXiv:2609.39139
Retrieval-augmented generation with an explicit belief state, so the system tracks how confident it should be as evidence arrives, conflicts, and goes stale — instead of treating every retrieval as equally authoritative.
Launder-Bench: Can Repository-Level Agentic Code Laundering Be Detected?
Under reviewICLR 2027
Coding agents can now rewrite a repository-scale codebase while preserving tested behaviour, which breaks the assumptions behind plagiarism detection, license auditing, and software provenance. Launder-Bench defines repository-level agentic code laundering and builds the first repository-scale provenance benchmark for it from 282 real-world repositories, keeping a rewrite only after execution-based verification under a source-hidden reconstruction protocol. 1,128 pairwise instances ask whether two repositories share provenance; 229 attribution instances ask which candidate is the source, with abstention allowed when evidence is insufficient. The strongest static detector reaches only 0.66 AUROC, and repository encoders rank the true source first on up to 22% of attribution queries while earning calibrated attribution credit on just 1–2%.
Self-Evolving Agents
Systems that improve from experience, published knowledge, and verified feedback without drifting away from their original specification.
Paper2LLM++: Continual Self-Evolution of Language Models From Research Papers
Under reviewTACL · arXiv:2610.02793
Treating the literature as a continual training signal: extracting methods from research papers and folding them back into the model, so capability accrues from reading rather than from retraining on scraped text.
SkillOps: Managing LLM Agent Skill Libraries as Self-Maintaining Software Ecosystems
Under reviewICLR 2027 · arXiv:2605.13716
An agent's skill library is a codebase, so it should be run like one — versioned, tested, deprecated, and repaired — which is what stops accumulated capability from quietly decaying.
World Models for Networks
Learned, action-conditioned environments for complex networks, used as the evaluation substrate for automated algorithm design.
Network World Models as Environments for Algorithm Design on Complex Systems
Under reviewICLR 2027 · arXiv:2610.01048
An action-conditioned Network World Model learns how diffusion unfolds under interventions — applying each action to the network, then predicting what follows — and serves as a fast evaluator inside an algorithm design loop, where a coding agent refines executable algorithms from full rollouts, action-level credit, and counterfactual probes. Across eight network tasks and five diffusion models, the designed algorithms match or exceed the strongest reported baseline in 138 of 141 settings, with rollouts up to 14.5× faster than Monte Carlo simulation.
Efficient Agent Infrastructure
The plumbing — profiling, routing, and orchestration that keep multi-model systems affordable at scale.