Teo Wiki References
A categorized study index for papers, techniques, and operating notes that Teo wants to preserve.
This page is the table of contents for Teo’s public study notes. Each paper or useful link gets its own page; this index only keeps the catalog, category map, and short summaries.
Category map
LLM foundations
Prompting and reasoning basics. Put papers here when they explain how LLM reasoning is elicited or evaluated before tool use.
Examples: Chain-of-Thought, self-consistency, reasoning evaluation, inference-time scaling.
RL for LLMs / Post-training
Reinforcement learning, reward modeling, preference optimization, and systems for post-training LLMs and agentic models.
Examples: PPO, GRPO, DAPO, reward models, PRMs, RLHF, agent RL, rollout systems.
Agent architectures
Patterns for agents that plan, use tools, call APIs, browse, write code, or interact with environments.
Examples: ReAct, tool use, planning loops, reflection, memory, multi-agent coordination.
Agent operations
Practical operations for running, separating, debugging, and maintaining agents in real workflows.
Examples: Hermes profiles, gateway operation, scheduled agents, deployment notes.
LLM foundations
Start here when studying the basic mechanisms that make LLMs reason or appear to reason.
Chain-of-Thought prompting
A compressed study note on CoT prompting: why intermediate reasoning tokens work as a text scratchpad, what changed in prompting, and how CoT connects to ReAct.
RL for LLMs / Post-training
Use this category for RL algorithms, reward modeling, RLHF/preference optimization, rollout systems, and task-specific agent RL.
InstructGPT and RLHF
A compressed study note on the SFT → Reward Model → PPO/PPO-ptx pipeline: how human preferences become a reward objective for instruction-following LLMs.
DeepSeekMath, PPO, and GRPO
A compressed study note on how DeepSeekMath uses CoT/PoT SFT and GRPO, and how policy, reward, value, reference, old-policy ratios, clipping, and KL penalties fit together in LLM RL.
RLVR and reasoning capacity beyond the base model
A compressed study note on whether RLVR/PPO/GRPO creates new reasoning capacity or mainly improves sampling efficiency for reasoning paths already present in the base model.
RL for LLMs: reading sprint map
A categorized study map built from the original RL-for-LLMs reading list. Teo is using that homepage as the main reference for this week’s RL paper sprint, with motivation, key techniques, and reading-depth rankings preserved here.
Agent architectures
Put papers here when they turn an LLM from a text generator into an agent that reasons, chooses actions, observes results, and revises plans.
ReAct: reasoning and acting in language models
A compressed study note on the ReAct pattern: why CoT-only and act-only are insufficient, how Thought/Action/Observation loops work, and where the implementation is fragile.
LangGraph + MCP agent study notes
A compressed note from Teo’s study of the braincrew-lab LangGraph MCP agent codebase: how Streamlit, MCP server configs, MultiServerMCPClient, ReAct agents, memory, and tool execution fit together.
Agent operations
Keep practical engineering notes here: setup, profiles, memory separation, gateways, cron jobs, and deployment patterns.
Hermes Telegram multi-bot profile isolation
A distilled note from a Korean field report about running multiple Telegram bots with Hermes. The key idea is that multi-bot operation is a profile and gateway isolation problem, not just a token-management problem.