I'm currently working on scaling autonomous discovery at Intology.

I graduated with a master's from the University of Illinois at Urbana-Champaign, supervised by Professor Lingming Zhang. My prior work has demonstrated state of the art performance on PostTrainBench, RE-Bench, MLE-Bench Lite, SWE-Bench, and SWE-Bench Multimodal (among other benchmarks) and has been used by OpenAI and DeepSeek in evaluating their models' software engineering capabilities.

News

  • August 3rd, 2026
    Released Scaling Automated Post-Training: Locus sets SOTA on PostTrainBench (44.7, 2.9 points ahead of the next frontier agent), reaches 51.6% on PostTrainBench+ surpassing the official human-tuned Qwen3-1.7B (49.4%), ranks 4th by peak average rank across all live prize-money Kaggle competitions, and ships a production Bubble model at ~2.8× lower error and 105× lower cost.
  • May 19th, 2026
    Led the release of NanoGPT-Bench, an autonomous AI R&D benchmark built on the NanoGPT Speedrun. Frontier coding agents (Claude Code, Codex, Autoresearch) given a 512 H100-hour budget recover less than 10% of 5 months of human world-record progress (top baseline 9.3%), spending the majority of their compute on hyperparameter tuning rather than the algorithmic research that drives 75.8% of human records.
  • November 19th, 2025
    Previewed Locus, the first AI system to outperform human experts at AI R&D - matching/beating METR's human expert baseline on RE-Bench (1.30 vs 1.27 over a continuous 64-hour run), setting SOTA on MLE-Bench Lite (medals in 77% of competitions vs prior 68%), and achieving SOTA on KernelBench
  • May 27th, 2025
    Zochi published in ACL 2025, “Tempest: Automatic Multi-Turn Jailbreaking of Large Language Models with Tree Search” - the first fully autonomous publication in a top scientific venue by an AI system.
  • March 14th, 2025
    Debuted Zochi, an artificial scientist which produced papers averaging 7.67/10 on NeurIPS guidelines, 3 of which were accepted to ICLR 2025 workshops.
  • February 26th, 2025
    Agentless Lite doubles state-of-the-art on SWE-bench Multimodal from 12.19% to 25.34% for 1/4th of the cost without even requiring a runtime environment!
  • February 14th, 2025
    Released Agentless Lite - a generalized, lightweight adaptation of the Agentless scaffold which is competitive with SOTA agents while only requiring sampling from a single RAG prompt.
  • January 31st, 2025
    Agentless used by both DeepSeek and OpenAI to evaluate their new reasoning models (r1 and o3-mini) on SWE-bench
  • Dec 2nd, 2024
    Integrated Agentless with Claude 3.5 Sonnet to achieve 40.7% solve rate on SWE-bench lite and 50.8% solve rate on SWE-bench verified
  • Oct 28th, 2024
    Released OpenAutoCoder-Agentless 1.5 which increases Agentless performance from 27.3% to 32.00% on SWE-bench lite
  • September 26, 2024
    MedCalc-Bench was accepted as an oral presentation for the NeurIPS 2024 Datasets and Benchmark Track
  • September 12, 2024
    Agentess was used by OpenAI as their scaffold of choice for evalauting gpt-4o, o1-mini, and o1-preview's model autonomy as part of their preparedness framework
  • July 1st, 2024
    Released OpenAutoCoder-Agentless 1.0! Agentless currently is the best open-source approach on SWE-bench lite with 82 fixes (27.3%) and costing on average $0.34 per issue.

Technical Reports

All Intology writing

  1. Automated model post-training

    Scaling Automated Post-Training

    August 2026

    Locus plans and steers many post-training experiments in parallel over multi-day horizons. It leads PostTrainBench, and with additional compute surpasses the official Qwen3-1.7B-Instruct checkpoint across the PostTrainBench+ suite. Locus has also post-trained a model deployed in production.

    Selected results

    • 44.7 PostTrainBench SOTA (verified)
    • 51.6 on PostTrainBench+, above Qwen3-1.7B-Instruct
    • 4th among accounts entered in all live prize-money Kaggle competitions

  2. Measuring real research ability

    NanoGPT-Bench

    May 2026

    Benchmarks for research agents are easy to saturate and easy to contaminate. NanoGPT-Bench drops agents into the GPT-2 pretraining speedrun at a fixed human world record and asks them to make it faster, with no internet and no human help. Frontier coding agents recover less than 10% of what the human community achieved over the following five months, spending most of their compute on hyperparameter tuning instead of the algorithmic work that actually moves the record.

    Selected results

    • 9.3% of human progress recovered, at best
    • 512 H100-hours per agent

  3. Artificial Scientist for AI R&D

    Previewing Locus

    November 2025

    Introducing Zochi's successor: a long-horizon research system that autonomously plans, executes, and learns from experiments over continuous multi-day runs. The initial Locus release matched or exceeded expert-human performance on RE-Bench and established new state of the art results across several AI R&D evaluations.

    Selected results

    • 1.30 on RE-Bench, above the 1.27 human-expert baseline
    • Medals in 77% of MLE-Bench Lite competitions
    • State of the art on KernelBench

  4. AI-generated discovery at a top venue

    Zochi at ACL 2025

    May 2025

    Zochi's paper Tempest was accepted to the main proceedings of ACL 2025 — the first fully AI-generated discovery to clear peer review at an A*-ranked scientific venue. Zochi independently developed the method, implemented it, ran the experiments, and wrote the paper, excluding figures and minor formatting fixes.

    Selected results

    • Accepted to the ACL 2025 main proceedings
    • Top 8.2% of submissions by meta-review score
    • 97% attack success rate on GPT-4

  5. The first Artificial Scientist

    The Zochi Technical Report

    March 2025

    Zochi takes a research question from literature review through experimentation to a written paper. The technical report describes the end-to-end system and its first autonomous research results, including multiple papers accepted to ICLR 2025 workshops.

    Selected results

    • Three papers accepted to ICLR 2025 workshops
    • 7.67/10 average score under NeurIPS review guidelines
    • End-to-end literature review, ideation, experimentation, and writing

Publications