Let's Verify Step by Step · HackerPapers discussions

Let's Verify Step by Step — mentioned by dilyevsky

Wed, 16 Sep 2026 18:11:32 GMT

there is a number of different methods. One is a separate model/classifier trained with human annotation. Here's a paper by OAI: https://arxiv.org/abs/2305.20050

Paper and discussion history. A mention is not an endorsement.

Let's Verify Step by Step — mentioned by highfrequency

Tue, 28 Jan 2025 17:53:40 GMT

First graph tells the story - below a certain model size (500m params), reinforcement learning is close to useless. Above this (task-dependent) model size threshold, reinforcement learning basically works. I suspect this is what we saw play out with math/coding reasoning models - until recently, the base models were not good enough for ~random output search to hit on a correct path with any reasonable frequency. Below this threshold of base model intelligence, the only efficient way forward was to collect plain supervised data (either through human labeled math problem solutions [1] or…

Paper and discussion history. A mention is not an endorsement.

Let's Verify Step by Step — HN submission

Fri, 13 Sep 2024 02:30:53 GMT

Let's Verify Step by Step

3 points; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.

Let's Verify Step by Step — mentioned by llmfan

Sat, 31 Aug 2024 22:44:32 GMT

"just a regular LLM except [trained on very different data]." I'm not saying there's some big moat, anyone can read https://arxiv.org/pdf/2305.20050, but not all synthetic data is created equal. Strawberry I'm sure generates beautiful, valid chain-of-thought reasoning data. Wouldn't surprise me if OpenAI is just significantly ahead of the competition.

Paper and discussion history. A mention is not an endorsement.

Let's Verify Step by Step — HN submission

Fri, 30 Aug 2024 17:21:02 GMT

Let's Verify Step by Step

2 points; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.

Let's Verify Step by Step — HN submission

Thu, 29 Aug 2024 12:57:54 GMT

Let's Verify Step by Step

1 point; 2 discussion replies on HN. Paper and source history. A submission is not an endorsement.

Let's Verify Step by Step — mentioned by rvz

Mon, 27 Nov 2023 13:31:02 GMT

Great, a worthy competition which being first and correctness completely matter and may the best AI model win. Hopefully DeepMind doesn’t enter otherwise they would smoke everyone else other than OpenAI. Speaking of which, I’ll give you a 1% chance of winning with this paper from OpenAI: [0] with the MATH dataset: [1] EDIT: Why the downvotes? I'm trying to help you here and give you a starting point to win the competition? What's wrong with helping others? [0] https://arxiv.org/abs/2305.20050 [1] https://github.com/hendrycks/math

Paper and discussion history. A mention is not an endorsement.

Let's Verify Step by Step — mentioned by ratg13

Sat, 25 Nov 2023 03:31:19 GMT

This is the one you want https://arxiv.org/pdf/2305.20050.pdf It’s different from the link the other user posted. This one is specifically about the Q* process

Paper and discussion history. A mention is not an endorsement.

Let's Verify Step by Step — HN submission

Sun, 04 Jun 2023 06:27:24 GMT

Let's Verify Step by Step

1 point; 1 discussion reply on HN. Paper and source history. A submission is not an endorsement.

Let's Verify Step by Step — HN submission

Sat, 03 Jun 2023 11:01:38 GMT

OpenAI: How to Train Reasoning in LLMs

2 points; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.