Interlace AI @interlaceIA

@MultiverseCompu Amazing job!

We are looking for collaborators who advocate for AI as a right. If you want to help us build and develop open-source models - DM Estamos buscando colaboradores que defienden que la IA sea un derecho. Si quieres ayudarnos a construir y desarrollar modelos open-soruce - DM

Model card, code, and every trajectory behind these numbers: huggingface.co/InterlaceAI/be… Technical note: doi.org/10.5281/zenodo…

Introducing Best-of-N 1.1.7 Inference-time compute for any language model. Sample a frozen model N times, select among the samples, get a better answer. No training, no fine-tuning, no new weights. Measured on 8 models. Same 200 GSM8K problems, same prompt, same settings, weights frozen: SmolLM2-1.7B 21.7% → 59.0% +37.3 Qwen2.5-0.5B 45.5% → 65.5% +20.0 Qwen2.5-1.5B 68.8% → 80.0% +11.2 Phi-3-mini 3.8B 80.9% → 91.0% +10.1 Qwen2.5-3B 83.0% → 91.0% +8.0 Qwen2.5-Math-1.5B 84.3% → 89.5% +5.2 Qwen2.5-7B 90.1% → 94.0% +3.9 Qwen3.8-27B 95.1% → 97.5% +2.4 Every one improved. Not one was trained. What you get: → Six selectors behind one call — majority, self-certainty, verifier, verifier-argmax, oracle, and random as the baseline you measure against → Coverage and accuracy reported separately, so you know whether to sample more or select better → Symbolic answer comparison: 1/2, 0.5 and \frac{1}{2} count as one vote → Abstention accounting — a truncated trajectory doesn't get to guess → Adapters for any published reward model, with the licence checked and the sigmoid applied for you → Shuffle the pool, same answer. Every selector, guaranteed. pip install "bestofn[math]" from bestofn import BestOfN engine = BestOfN("your/model", n=32) r = engine.solve(problem) r.answer, r.covered(gold) Runs on anything torch runs on. vLLM backend for large N. All 25,600 trajectories published in full — the complete reasoning text of every sample, not a summary. Reproduce every figure in two minutes on a laptop, no GPU. Apache 2.0. Free for research and commercial use.

Best-of-N is on PyPI. pip install "bestofn[math]" from bestofn import BestOfN engine = BestOfN("Qwen/Qwen2.5-0.5B-Instruct", n=16) r = engine.solve(problem) r.answer what it returns r.covered(gold) whether it was reachable at all The core has no dependencies. The [math] extra adds symbolic answer comparison, so 1/2, 0.5 and \frac{1}{2} count as one vote instead of three. Apache 2.0. pypi.org/project/bestof…

Works with any causal LM, at any N: BestOfN("Qwen/Qwen2.5-0.5B-Instruct", n=16) BestOfN("/path/to/your/local/model", n=128) Bring your own reward model too — the adapters handle the sigmoid and check the licence for you. Everything here 👇 huggingface.co/InterlaceAI/be…

Which means you do not have to believe any of this. pip install "bestofn[math]" python scripts/analyse.py Every number in this thread comes back out in about two minutes. On a laptop. No GPU. Apache 2.0. Free forever, commercial use included.

Introducing Best-of-N. Your language model already knows more than it tells you. We took a 0.5B model, changed nothing about it, and moved it from 45% to 66.5% on GSM8K. No training. No fine-tuning. No new weights. Here is how. 🧵

Being clear about what's mine and what isn't: Best-of-N and self-consistency are established techniques (Cobbe et al. 2021, Wang et al. 2022). I didn't invent them. What I built is an open, tested implementation — 45 tests, every raw per-problem output published, and a script that replays them so you can check the numbers instead of believing me. 📄 Full report: doi.org/10.5281/zenodo… This is v1: maths and short-answer reasoning. If there's real interest, v2 goes after the harder target — agentic tasks and demanding code generation, where a verifier can actually RUN the code instead of guessing whether it works. That turns a 0.91 verifier into a near-perfect one. Tell me what you'd want it to handle.

So I trained a verifier: a second model whose only job is to read a solution and judge whether it's correct. 0.910 ROC-AUC. Instead of counting votes, it weights them by confidence. A confident minority can now overturn an unconfident majority. majority vote 35.6% + verifier 52.2% (+16.6 points, 90 AIME problems, N=32) It works outside maths too, same frozen model, no retraining: GPQA 33.8 → 43.4 · GSM8K 87.2 → 92.8 Three lines to use it with any model on the Hub: from bestofn import BestOfN engine = BestOfN("your/model", n=32) engine.solve(problem).answer 📦 huggingface.co/InterlaceAI/be…

@AndrewBolis AI