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

Implementation of paper: Flux Already Knows – Activating Subject-Driven Image Generation without Training

[ICML 2026] Decoding Tree Sketching (DTS): a training-free & model agonistic & plug-in framework for LLM parallel reasoning.

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

Efficient Test-Time Scaling for Small Vision-Language Models, official implementation of the ICLR'26 paper, test-time scaling via test-time augmentation

The official implementation for paper "AdaExplore: Failure-Driven Adaptation and Diversity-Preserving Search for Efficient Kernel Generation"

[NeurIPS 2025] Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search

[CVPR2026 Findings] VHS: Verifier on Hidden States, an efficient inference-time scaling verification framework for DiT-based image generation.

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.

[COLM 2025] Adaptive Skill-based In-context Math Instruction for Small Language Models

[ECCV'26] LatSearch: Latent Reward-Guided Search for Faster Inference-Time Scaling in Video Diffusion

Open-source research engineering project for building the end-to-end post-training stack for reasoning language models, including SFT, preference learning, RLHF/RLVR, evaluation, inference-time scaling, and scalable systems for frontier-level reasoning.

Experiments on System 2 reasoning — neuro-symbolic learning, inference-time scaling, LLM agents, RL post-training in LLMs, and Graph-based retrieval.

How Chain-of-Thought Budgets Induce Overconfidence in LLMs. Investigating Calibration Drift Under Reasoning (CDUR), hypothesis lock-in mechanisms, and dynamic token budgeting via the CABStop optimal stopping algorithm.

Two agent skills to keep a long-running AI agent from getting stuck: north-star (delivery-point self-correction hook) + moa (N+1 multi-model consultation).

Chain-of-density study of 'Recursive Language Models' (Zhang, Kraska & Khattab 2025, MIT CSAIL, arXiv:2512.24601) - five-tier note, locator-verified. The prompt-as-environment paradigm Trellis implements: REPL offloading, recursive sub-calls, 10M+ token inputs.

Empirical study of latent quality in LLMs under critical engagement

Benchmark of inference-time scaling strategies for LLM reasoning — chain-of-thought, repeated sampling, and self-consistency across accuracy, latency, and compute efficiency.

Official code for "Student Guides Teacher: Weak-to-Strong Inference via Spectral Orthogonal Exploration" (SOE, ACL 2026 Oral). A training-free, test-time scaling method that fixes reasoning collapse in LLM mathematical reasoning via orthogonal probing.

Code for "Challenges in Inference Time Scaling with Uncertainty Aware Tree Search" @ ICLR 26 Workshops