LIMO: Less is More for Reasoning · HackerPapers discussions

LIMO: Less is More for Reasoning — mentioned by sebzim4500

Tue, 11 Feb 2025 21:01:36 GMT

Yes, but in this case the RL stage doesn't actually need a huge amount of data. Most of the data is ingested in the pre-training phase, which in this case was done by Alibaba. There was a recent paper [1] that showed that for one particular model it was possible to get very good reasoning performance by finetuning on under a thousand examples of reasoning. [1] https://arxiv.org/abs/2502.03387

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

LIMO: Less is More for Reasoning — mentioned by shashanoid

Sun, 09 Feb 2025 19:03:51 GMT

Love prepending 'explain' to arxiv links these days xD https://explainarxiv.org/abs/2502.03387

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

LIMO: Less is More for Reasoning — HN submission

Sun, 09 Feb 2025 16:33:28 GMT

LIMO: Less Is More for Reasoning

389 points; 128 discussion replies on HN. Paper and source history. A submission is not an endorsement.

LIMO: Less is More for Reasoning — HN submission

Thu, 06 Feb 2025 11:37:59 GMT

LIMO: Less Is More for Reasoning

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