LLM Agent

A paper proposes a new approach called "Cache Engineering": letting the agent directly manipulate the active KV cache without modifying the trajectory or re-prefilling.

The Bellman Language Model (BLM) models repeated decisions as a Markov decision process (MDP), solves it once and reuses the solution, to improve the efficiency of LLM agents.

In agent workflows of limited scope, the decision space gradually saturates, and the agent repeatedly encounters the same states.

Agents need to monitor their own performance and proactively discard outdated knowledge when performance degradation is detected.

Current LLM agents often reason effectively once valid evidence is obtained, but they struggle to actively acquire the right evidence and to recognize when previously acquired evidence is no longer valid.

Under implicit rule drift, latent rules change without notification, and the agent must detect from performance degradation that its prior knowledge is outdated and re-explore.

LLM agents may, due to language priors, persistently output actions that violate preconditions with high confidence, thereby falling into ineffective loops.

Experiments on frontier LLM agents show that, compared with standard simulation-augmented baselines, FunCAD-Harness significantly improves functional success rates.

The agent framework achieved a 77.4% end-to-end autonomous reproduction rate in the SEER-Repro experiment.

A paper published on arXiv, "AI-Driven Reproducibility and Analytical Robustness in Scientific Discovery" (arXiv:2610.23845), proposes a framework for solving analytical reproducibility problems at scale using autonomous LLM agents.

The agent framework identified analytical fragility in 9.5% of the reproducible findings.

Never execute agent-generated analyses on the host runner; strict resource cgroups, disabled network egress, and deterministic random seeds should be enforced.

AgentOnto is an ontology-guided LLM agent framework that improves the reliability and adaptability of enterprise data analysis by constructing, using and updating an enterprise knowledge ontology.

Traditional LLM agents lack an explicit representation of an enterprise's scattered knowledge, making them prone to erroneous results and unable to effectively absorb feedback.

Explicit knowledge representation is a key path to improving the reliability of LLM agents.

The developed AI agent successfully reproduced 77.4% of the atomic claims on the SEER-Repro benchmark.

Limitations of the current method include that the reproduction rate is not 100%, and that robustness analysis depends on accurate identification of analytical choices, which may be limited by the level of detail in the paper's methods section.

LLM agents can read methods sections, generate and debug analysis code, and iteratively correct errors when they occur, thereby compressing manual reproduction workflows into batch-executable pipelines.

Agent success depends heavily on the granularity of the paper's methods section and on whether the data can be accessed programmatically.

A multi-agent pipeline refers to breaking a task down among multiple independently running LLM agents, each with its own prompt, role, or perspective, whose outputs are then aggregated, opposed, or cross-validated; common forms include fan-out, review panels, and adversarial validation.