Deep Think with Confidence · HackerPapers discussions

Deep Think with Confidence — mentioned by NitpickLawyer

Sun, 19 Oct 2025 14:50:28 GMT

> They really can’t. Token prediction based on context does not reason. Debating about "reasoning" or not is not fruitful, IMO. It's an endless debate that can go anywhere and nowhere in particular. I try to look at results: https://arxiv.org/pdf/2508.15260 Abstract: > Large Language Models (LLMs) have shown great potential in reasoning tasks through test-time scaling methods like self-consistency with majority voting. However, this approach often leads to diminishing returns in accuracy and high computational overhead. To address these challenges, we introduce Deep Think with Confidence…

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

Deep Think with Confidence — mentioned by NitpickLawyer

Sun, 19 Oct 2025 14:49:33 GMT

AIME was saturated with tool use (i.e. 99%) for SotA models, but pure NL, no tool still perform "unreasonably well" on the task. Not 100% but still within 90%. And with lots of compute it can reach 99% as well, apparently [1] (@512 rollouts, but still) [1] - https://arxiv.org/pdf/2508.15260

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

Deep Think with Confidence — mentioned by swores

Sun, 24 Aug 2025 17:56:35 GMT

Copied from the paper (halfway down page 6: https://arxiv.org/pdf/2508.15260 ) > "Specifically, DeepConf-low uses top η= 10% (corresponding to the 90th percentile) and DeepConf-high uses top η = 90% (corresponding to the 10th percentile) uniformly across all settings. This threshold ensures that during online generation, traces are terminated when their confidence falls below the level that retains the top η% highest-confidence traces from the warmup phase." I'm not sure if I'm parsing it right, but are they using "low" and "high" as descriptors of the number used as the %, meaning that the…

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