Advanced Large Language Model Agents

Author

Chao Ma

Published

August 25, 2026

Notes on advanced LLM agents, with an emphasis on reasoning, search, evaluation, feedback, and self-improvement.

An Agent’s Reasoning Should Not Grant Its Permissions Berkeley Lecture 12: prompt injection, end-to-end risk evaluation, and the scope of deterministic privilege control.

A Stronger Statement Can Make the Proof Easier Berkeley Lecture 11: generalized compiler contracts, COPRA feedback, and concept-guided scientific discovery with LaSR.

A Checked Proof Still Needs a Plan and a Library Berkeley Lecture 10: how thoughts, proof sketches, premise selection, and project context help find checker-accepted proofs.

Autoformalization and Theorem Proving Why a checked proof can answer the wrong question: translation fidelity, premise retrieval, and hidden diagrammatic assumptions.

AlphaProof: Reinforcement Learning Meets Formal Mathematics Lean verification, proof search, and test-time learning from related problems.

Multimodal Agents: From Perception to Action OSWorld evaluation, executed training trajectories, and the connection between grounding and planning.

Multimodal Autonomous AI Agents Visual grounding, tree search, and training from filtered web-agent trajectories.

Coding Agents and AI for Vulnerability Detection Evaluation, control flow, tool interfaces, and Big Sleep’s loop for testing vulnerability hypotheses.

Open Training Recipes: LLM Reasoning Tulu 3’s post-training pipeline, verifiable rewards, s1 test-time scaling, and OLMo mid-training.

Reasoning, Memory, and Planning of Language Agents HippoRAG and associative memory, grokking and implicit reasoning, and WebDreamer’s world models for planning.

Learning to Self-Improve and Reason with LLMs System 1 and System 2 reasoning, self-rewarding training, verifiable rewards, Thought Preference Optimization, Meta-Rewarding, and reasoning-based evaluators.

Inference-Time Techniques for LLM Reasoning Prompting strategies, candidate sampling and selection, Tree of Thoughts, voting, Reflexion, Self-Refine, and self-debugging.

--- title: "Advanced Large Language Model Agents" author: "Chao Ma" date: "2026-08-25" --- Notes on advanced LLM agents, with an emphasis on reasoning, search, evaluation, feedback, and self-improvement. --- ::: {.content-grid} ::: {.content-card} **[An Agent's Reasoning Should Not Grant Its Permissions](safe-and-secure-agentic-ai.qmd)** Berkeley Lecture 12: prompt injection, end-to-end risk evaluation, and the scope of deterministic privilege control. ::: ::: {.content-card} **[A Stronger Statement Can Make the Proof Easier](abstraction-and-discovery.qmd)** Berkeley Lecture 11: generalized compiler contracts, COPRA feedback, and concept-guided scientific discovery with LaSR. ::: ::: {.content-card} **[A Checked Proof Still Needs a Plan and a Library](advanced-topics-theorem-proving.qmd)** Berkeley Lecture 10: how thoughts, proof sketches, premise selection, and project context help find checker-accepted proofs. ::: ::: {.content-card} **[Autoformalization and Theorem Proving](language-models-autoformalization-theorem-proving.qmd)** Why a checked proof can answer the wrong question: translation fidelity, premise retrieval, and hidden diagrammatic assumptions. ::: ::: {.content-card} **[AlphaProof: Reinforcement Learning Meets Formal Mathematics](alphaproof-reinforcement-learning-formal-mathematics.qmd)** Lean verification, proof search, and test-time learning from related problems. ::: ::: {.content-card} **[Multimodal Agents: From Perception to Action](multimodal-agents-perception-to-action.qmd)** OSWorld evaluation, executed training trajectories, and the connection between grounding and planning. ::: ::: {.content-card} **[Multimodal Autonomous AI Agents](multimodal-autonomous-ai-agents.qmd)** Visual grounding, tree search, and training from filtered web-agent trajectories. ::: ::: {.content-card} **[Coding Agents and AI for Vulnerability Detection](coding-agents-vulnerability-detection.qmd)** Evaluation, control flow, tool interfaces, and Big Sleep’s loop for testing vulnerability hypotheses. ::: ::: {.content-card} **[Open Training Recipes: LLM Reasoning](open-training-recipes-llm-reasoning.qmd)** Tulu 3's post-training pipeline, verifiable rewards, s1 test-time scaling, and OLMo mid-training. ::: ::: {.content-card} **[Reasoning, Memory, and Planning of Language Agents](reasoning-memory-planning-language-agents.qmd)** HippoRAG and associative memory, grokking and implicit reasoning, and WebDreamer's world models for planning. ::: ::: {.content-card} **[Learning to Self-Improve and Reason with LLMs](learning-to-self-improve-and-reason-with-llms.qmd)** System 1 and System 2 reasoning, self-rewarding training, verifiable rewards, Thought Preference Optimization, Meta-Rewarding, and reasoning-based evaluators. ::: ::: {.content-card} **[Inference-Time Techniques for LLM Reasoning](inference-time-techniques-for-llm-reasoning.qmd)** Prompting strategies, candidate sampling and selection, Tree of Thoughts, voting, Reflexion, Self-Refine, and self-debugging. ::: :::