Showing 1-12 of 15
Nine lessons implement a minimal agent loop—tool calls, memory, state, policy gates, self-scheduling—in about 60 lines of Python. Run in-browser via Pyodide with mock or Groq LLM.
Automate translation of Pokémon GBA ROM hacks with Meowth: extract text, use LLMs to translate while preserving in-game codes and fonts, then rebuild a playable ROM via GUI or CLI.
A commit-gated pipeline (Prompt→LLM→Validation→Normalizer→Retry→Commit) keeps AI-generated structured files off disk until they pass schema checks, revealing schema gaps.
Build an AI-chat evaluation harness for He Xin’s PEPC formal language to test expressiveness, contradiction handling, and alignment with wargame baselines — includes artifacts and metrics.
Simulate conversations across Retell, VAPI, LiveKit and Bland with Voicetest. It normalizes configs via an AgentGraph IR, scores turns with LLM judges, and adds compliance checks.
Use Hallucinating Splines' REST API to run an LLM as a city mayor in a live Micropolis simulation. See how agents manage zoning, roads and power and why spatial planning trips them up.
Hands-on prototype: build a Cart Assistant that uses text or photo shopping lists, OCR, and user order history to prefill Uber Eats grocery carts for review before checkout.
Describes PCE (Planner-Composer-Evaluator) that turns LLM reasoning assumptions into decision trees, then scores paths by likelihood, goal gain and execution cost to reduce communication.
Agent-Omit trains LLM agents to omit redundant internal thoughts and observations using cold-start omission data plus omit-aware RL; includes a KL-divergence bound and Agent-Omit-8B results.
AgentArk distills multi-agent debate into a single LLM via three hierarchical distillation strategies, shifting computation to training to cut inference cost while preserving reasoning.
ORBIT trains LLMs via cross-episode meta-RL so models learn from interaction traces at inference; authors report Qwen3-14B matches GPT-5.2 on unseen environments after meta-training.
Guide to building Scalable Interactive Oversight: decompose intent into a decision tree, collect low‑burden node feedback, aggregate signals, and gain 54% better alignment.