test-time-compute

Implement a reasoning LLM in PyTorch from scratch, step by step

Repo for AI Agents The Definitive Guide

Inference-time scaling for LLMs-as-a-judge.

Multi-agent demo platform for Titans (arXiv:2501.00663) — neural networks that learn to memorize at test time. 7 AI agents, native desktop UI.

Official repository of the spotlight ICML 2025 paper, PokeChamp: an Expert-level Minimax Language Agent.

Compound model panel for pi: parallel model calls, then one synthesis response.

Airgapped closed-corpus QA loop: a self-hosted Qwen3.6 agent explores a .zip dataroom under a token budget with local tools

Multi-Agent Verification: Scaling Test-Time Compute with Multiple Verifiers

[NeurIPS 2025 Spotlight] Official implementation of "Learning with Calibration: Exploring Test-Time Computing of Spatio-Temporal Forecasting"

Evolving agent harnesses: a research program on how far N orchestrated calls of a small model can rival a frontier model. We evolve the harness (structure + prompts) with reflective optimizers + a verified-acceptance gate.

Test-Time Memory Framework: Control Hallucinations in Foundation Models

Tree execution engine for LLM inference: fork, merge, prune KV cache at token granularity

Official implementation of Dynamic Parallel Tree Search for accelerating LLM reasoning with test-time parallel search.

Code for the paper "Specification Self-Correction: Mitigating In-Context Reward Hacking Through Test-Time Refinement"

Turn LLM coding agents (Claude Code, Codex) from next-token predictors into divergent thinkers. A research-grounded cognitive engine + drop-in Claude Code skills for creativity, reasoning & robustness, with 6 honest reproducible benchmarks and 137 verified papers.

Code for ICML 2025 How Do Large Language Monkeys Get Their Power (Laws)?

An experimental project using MCTS to refine LLM responses for better accuracy and decision-making.

Do stacked LLM verifiers actually multiply reliability? An empirical-Bayes inversion that recovers verifier correlation (ρ_v) and the reliability ceiling from accept/reject logs — and falsifies the independence "Odds Law."

Frontier-level mathematical reasoning from a 1.5B model: 83.3% on AIME 2024 via inference-time compute and learned verification. Technical report TR-2026-01.

Controlled experiments on recurrent-depth (looped) transformers — length extrapolation, adaptive test-time compute, and honest sub-1B negatives.