Research Blog

Deep dives into machine learning research papers, mathematical derivations, and implementation details.

Essential PyTorch: the 5% that shows up in almost every serious model build

December 1, 2025•15 min read

PyTorch looks huge from the outside. In practice, I build Transformers, CLIP, diffusion, flows, and VAEs using a small set of tensor and module patterns over and over. This post is that core.

Hierarchical Reasoning Model (HRM): recursion, deep supervision, and why ACT is challenging

December 14, 2025•12 min read

A deep dive into HRM architecture—recursive refinement, deep supervision, DEQ-style gradients, and a mechanical breakdown of ACT's halt/continue logic.

Tiny Recursive Model (TRM)

December 15, 2025•10 min read

TRM simplifies HRM by removing fixed-point theory, eliminating ACT's extra forward pass, and backpropagating through full unrolled recursion with no-grad refinement passes.

Designing a Deterministic LLM Agent

December 21, 2025•12 min read

A coding agent is a stochastic model wrapped in a deterministic scaffold. This post walks through how to build that scaffold with a ledger-style History, derived summaries, allowlisted tools, driver-owned tests, and reflection triggers.

December 31, 2025•10 min read

JanusFlow unifies image understanding and generation in a single LLM backbone using rectified flow. This post explains the architecture, training stages, representation alignment, and why the shared backbone matters.

LLM Repo Agent v2: What's changed since v1

January 1, 2026•8 min read

V2 brings Chat Completions adapters, JSON tool protocol, multi-turn message lists, prompt hardening with invariants, sandboxing, multithreading, Monte Carlo evaluation, and SFT/DPO fine-tuning pipelines.

Bug-fixing agent: json tool calling SFT

January 2, 2026•12 min read

SFT plan to instruction-tune Qwen2.5-7B for JSON tool-protocol compliance. Covers step-level dataset extraction from teacher traces, data quality filtering, QuixBugs train/test splits, and end-to-end pipeline commands.

Modern post-training part 1: PPO—the ancestor algorithm

January 26, 2026•15 min read

Part 1 of a 4-part series on modern post-training. PPO-RLHF is the ancestor algorithm—three models, on-policy rollouts, and a critic for variance reduction. Understanding PPO explains what DPO, GRPO, and GDPO delete and why.

Modern post-training part 2: Direct Preference Optimization (DPO)

January 30, 2026•12 min read

Part 2 of the modern post-training series. DPO removes the RL machinery from PPO-RLHF—no critic, no advantage estimation, no rollouts—while preserving KL-regularized alignment through a closed-form supervised objective on preference pairs.

Modern LLM post-training part 3: GRPO

January 30, 2026•15 min read

Part 3 of the modern post-training series. GRPO keeps PPO's optimizer but removes the critic—advantages are estimated from a group of rollouts per prompt instead of learning a value function.

Modern LLM post-training part 4: GDPO

February 2, 2026•12 min read

Part 4 of the modern post-training series. GDPO fixes GRPO's reward signal collapse in multi-reward setups by normalizing each reward dimension separately before aggregation.