Modelpublic

FINAL-Bench/Darwin-27B-RSI

sourceHugging Faceapache-2.0updated 11d agoView on Hugging Face

33likes1.2kdownloads

Model Card

Darwin-27B-RSI: A Model That Improved Itself — Learning Only From Its Own Solutions

Qwen3.5-27B family · 27B dense · Thinking mode · BF16 · Apache 2.0 No human-written answers. The model generated its own learning signal — and got measurably better.

Abstract

Darwin-27B-RSI is Darwin-27B-Opus after Recursive Self-Improvement (RSI): the model was improved using only signal it produced itself. During self-improvement, the model itself (its weights) improves by learning only from its own solutions. No human-written solutions or reasoning traces are used; correctness is checked automatically (agreement across its own samples and code execution).

Under an identical evaluation protocol, Darwin-27B-RSI improves over its parent on graduate-level science reasoning — +5.24 points on GPQA Diamond (single sample) and +3.79 points with majority voting — with every gain statistically significant in paired tests.

As the reasoning engine of Darwin-27B-JEV on the Decision Index, it lifts the hardest reasoning decisions: GPQA Diamond skill 0.31 → 0.71, GSM8K 0.61 → 0.97, MMLU-Pro 0.60 → 0.82.

Model-level RSI vs. harness-level RSI

Darwin-27B-RSI is Model-level RSI: the model itself (its weights) improves by learning only from its own solutions. No human-written solutions or reasoning traces are used; correctness is checked automatically. You download a new model file, and it is smarter on its own.

Harness-level RSI (for example, Google's RRSI) improves the prompts, tools and workflow around a fixed model — like rewriting an employee's manual, while Model-level RSI is the employee getting smarter. The two are complementary: a harness-level loop can run on top of a Model-level RSI model.

What Is RSI?

Most models improve only when people write more answers for them. Recursive Self-Improvement removes that bottleneck: the model works on problems, judges its own work, and learns from what it produced — then repeats. Each improved model becomes the starting point for the next improvement.

Darwin-27B-RSI demonstrates this loop on a 27B model:

—Human-written solutions or reasoning traces used: 0
—Direction of change: measurably better on held-out graduate-level science
—Contamination check: training problems share 0 items with the evaluation sets reported here

The training procedure itself is not released.

Results

Science reasoning (same protocol for both models)

Benchmark	Darwin-27B-Opus	Darwin-27B-RSI	Δ
GPQA Diamond (1 sample)	72.85	78.09	+5.24
GPQA Diamond (majority@16)	79.80	83.59	+3.79
SuperGPQA (1 sample)			+4.03

Both models were measured under the same protocol (single sample, identical sampling settings and token budget), so numbers differ from the Darwin-27B-Opus card, which reports a different protocol. All gains are statistically significant in paired tests.

Decision Index — as the reasoning engine of Darwin-27B-JEV

The Decision Index scores typed-decision engines on 43 benchmarks and ~121K decisions (chance-corrected: 0 = random, 1 = perfect). Darwin-27B-RSI handles the decisions that need real thinking:

Benchmark (skill)	before	with Darwin-27B-RSI
GPQA Diamond ★	0.31	0.71
GSM8K	0.61	0.97
CRUXEval	0.61	0.87
MMLU-Pro ★	0.60	0.82
BBH ★	0.68	0.83
CLadder	0.49	0.70

★ = gold benchmark (weighted 1.2× on the board). Darwin-27B-JEV: ≈ 61.1 under the v0.2.1 board rules (our recomputation; official score pending review). Full run: FINAL-Bench/Darwin-27B-JEV-decision-index.

Usage

Darwin-27B-RSI is a thinking model. Give it room to reason and read the answer after the reasoning block.

Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "FINAL-Bench/Darwin-27B-RSI"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")

messages = [{"role": "user", "content": "A ball is thrown upward at 40 m/s. For how long is it above 40 m? (g = 10 m/s²) Think, then give the final answer."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=8192, temperature=0.6, top_p=0.95, do_sample=True)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

vLLM

bash

vllm serve FINAL-Bench/Darwin-27B-RSI --max-model-len 32768

Recommended: temperature 0.6, top_p 0.95, generous token budget (8K–16K) for hard problems.

Model Details


Parent	FINAL-Bench/Darwin-27B-Opus
Architecture	Qwen3.5 family, 27B dense
Precision	BF16
Improvement method	Recursive Self-Improvement, no human labels
License	Apache 2.0
Developer	VIDRAFT · FINAL-Bench

Limitations and Disclosure

—Gains were measured on graduate-level science; other domains may change less.
—As a thinking model, it can produce long reasoning; cap max_new_tokens for latency-sensitive use.
—31 training problems (0.22% of the benchmark) overlap with the Decision Index MMLU set; the model learned only from its own solutions to them.
—Not affiliated with TypeSafe AI or its Jev product.

Citation

bibtex

@misc{darwin27b_rsi_2026,
  title  = {Darwin-27B-RSI: Recursive Self-Improvement without Human Labels},
  author = {VIDRAFT and FINAL-Bench},
  year   = {2026},
  url    = {https://huggingface.co/FINAL-Bench/Darwin-27B-RSI}
}