Humanfia at IBO 2024

Note

This is part of RSI Effort at NVIDIA Research. Humanize is an open agent loop/flow framework that led by NVIDIA Research, UCLA PolyArch, and MIT HAN Lab. We are skying the limit with the power of agents with community members.

The Humanfia team have aced all 100/100 theory tasks in the International Biology Olympiad (IBO) 2024 Theoretical Exam using a fully agentic, YOLO-style approach. All 400/400 statement verdicts match the official answer key, giving 100% agreement on the published grading target. According to IBO committe policy, the problems will be hold for two years. We will update on 2026 problem set once it becomes public.

We build with open source, and build for open source. We release everything including:

• the complete natural-language worked solutions for Theory A and Theory B;
• the official-answer summary and complete grading evidence;
• the task plan, worker assignments, and validation scripts used for problem solving.

As Jensen shared, We all love open models X open harness 🎉 and the combination achieves full score at every competiton:

• IMO2026 / IOI2026 / IPhO2026 / ICho2026 / IBO2024

Results

The Humanfia (GPT-5.6) workers solved all 100 Theory A and Theory B tasks. The results below report exact agreement with the answer sections embedded in the official English exam papers.

Part	Humanfia (GPT-5.6)	Official verdict agreement
Theory A	✅ 50/50 tasks	✅ 200/200 (100%)
Theory B	✅ 50/50 tasks	✅ 200/200 (100%)
Total	✅ 100/100 tasks	✅ 400/400 (100%)

The extraction and format error count is 0. The complete grading evidence is in GRADING.md.

This is an answer-key agreement grade for the 100 theory tasks, not an official Olympiad points calculation. The repository intentionally excludes every practical exam. The natural-language explanations were reviewed separately against the local official questions and are protected by the reviewed-solution manifest.

Quick start

Launch with the reproduction prompt

The complete launch prompt is checked in as plan.md. AGENTS.md supplies the offline worker policy. The launch shell creates a clean workspace containing the prompt, policy, and official PDFs, but none of the published solutions:

bash scripts/run-ibo2024.sh \
  --prompt plan.md \
  --model gpt-5.6-sol

The shell passes plan.md directly to codex exec. The prompt assigns one distinct worker to each of the 100 Theory A/B tasks and runs the review and assembly workflow. Add --dry-run to validate the inputs without creating a workspace or launching Codex.

Reproduce the public result

Place the two official English exam PDFs at these exact local paths:

• source/IBO2024 Theory A.pdf
• source/IBO2024 Theory B.pdf

Create the pinned local text extracts used by the offline validator:

mkdir -p source
pdftotext -layout "source/IBO2024 Theory A.pdf" source/part-a.txt
pdftotext -layout "source/IBO2024 Theory B.pdf" source/part-b.txt

Then regenerate the derived collection, check that committed outputs are current, and run the complete offline validation gate:

python3 scripts/build.py
python3 scripts/build.py --check
python3 scripts/validate.py
python3 scripts/test_validate.py
python3 -m py_compile scripts/build.py scripts/validate.py scripts/test_validate.py
git diff --check

The validator compares all 400 declared verdicts with the answer sections embedded in the official PDFs. It invokes only local tools and files: it does not browse, query a network service, or consult a third-party answer key.

The one-worker assignment contract in plan.md, the canonical coordinates in WORKERS.md, and the validation commands above define how the 100 isolated results are assembled and accepted.

Reproduce with Kimi-K3

The published solution corpus was produced by distinct gpt-5.6-sol:max workers and reviewed with Humanize RLCR, as recorded in WORKERS.md. This repository does not label those files as Kimi-generated.

Kimi-K3 uses the same Codex harness, task contract, offline boundary, and review loop as the archived GPT-5.6 runs. No separate Kimi launcher or Kimi CLI is needed. In the existing Codex harness configuration, make only these two changes:

• replace the API key in the Codex auth file with the Kimi API key;
• replace the Codex model name with the Kimi model name.

Then run the same checked-in prompt through the same Codex harness:

CODEX_HOME=/path/to/codex-home-with-kimi-api-key \
bash scripts/run-ibo2024.sh \
  --prompt plan.md \
  --model your-kimi-model-name

Keep the 100 task assignments, local official PDFs, prompts, output paths, and validation commands unchanged. Each worker still receives one coordinate from WORKERS.md, follows AGENTS.md, and is accepted only after the repository-wide validation passes.

Read the collection

• Official answer summary
• Theory Part A consolidated solutions
• Theory Part B consolidated solutions
• Part A individual solutions
• Part B individual solutions
• One-worker assignment ledger

Offline methodology and ownership

Exactly one distinct gpt-5.6-sol:max solver worker was assigned to each task, as recorded in WORKERS.md. Every solver worked offline, inspected the corresponding question and embedded official answer in the local PDF, and wrote one Markdown file. Humanize RLCR reviewers also work offline from the same local official sources.

The explicit official verdict controls each answer pattern. If the printed graph, a label, a unit, or the official prose appears inconsistent, the solution retains the official T/F value and adds a clearly labeled discrepancy note instead of silently changing the key. The prose explanations are original educational derivations, not copies of the embedded explanations.

The reviewed corpus is locked by reviewed-solutions.sha256. The invariant is: Do not update a hash merely to make validation pass. If an individual solution changes, first compare that complete file again with its local official question and embedded answer; only after that direct review may its manifest hash be regenerated.

Repository structure

• solutions/part-a/ and solutions/part-b/: the 100 independently authored source solutions.
• official-answers.tsv: official A-D patterns derived from the embedded answer sections.
• task-map.tsv: canonical source-grounded topic identity for every coordinate.
• reviewed-solutions.sha256: integrity manifest for the source-reviewed solution corpus.
• answers.md: generated 100-row answer and navigation index.
• theory-a-solutions.md and theory-b-solutions.md: generated ordered volumes.
• scripts/build.py: deterministic collection builder and freshness check.
• scripts/run-ibo2024.sh: isolated Codex launcher for the checked-in reproduction prompt.
• scripts/validate.py: end-to-end offline source, ownership, content, document, and navigation validator.
• scripts/test_validate.py: isolated negative fixtures for the validation contract.