Scaling Up Good Synthetic Reasoning We're hiring! ➡️ https://xitter.cf/t.co/Gqtk6KrIy6 / DM 💬 @NathanThinks
- Tweets142
- Following47
- Followers14.6K
- Likes289
ALT Pareto efficiency analysis reveals how models distribute computational resources across problems of varying difficulty. Cumulative token allocation curves for problems ordered from easiest to hardest, aggregated across MATH-500, OlympiadBench, and AIME datasets. Shaded regions indicate easy (0-50%) and hard (80-100%) problem ranges. Just Enough Thinking: Efficient Reasoning with Adaptive Length Penalties Reinforcement Learning https://arxiv.org/pdf/2506.05256
ALT Pareto efficiency analysis reveals how models distribute computational resources across problems of varying difficulty. Cumulative token allocation curves for problems ordered from easiest to hardest, aggregated across MATH-500, OlympiadBench, and AIME datasets. Shaded regions indicate easy (0-50%) and hard (80-100%) problem ranges. Just Enough Thinking: Efficient Reasoning with Adaptive Length Penalties Reinforcement Learning https://arxiv.org/pdf/2506.05256
ALT Related Work Section
ALT Related Work Section
ALT In many applications, such as science for example, the reasoning chain can be even more important than the final answer. Qualitatively, current models with “advanced” reasoning still struggle and produce many CoTs which are inconsistent or unfaithful, even when they obtain the correct final answer. Interestingly, even the single math example provided along with the announcement of the O1 model has an incomplete solution as it makes an unsubstantiated assumption on the form of the polynomial ℎ(𝑥) = 𝑥 2 − 𝑐 without proving that the coefficient of 𝑥 is in fact zero (OpenAI, 2024). Under a competitive setting this would likely receive 5/7 points, even though the final answer is correct. It is worth noting that the model actually verifies and proves this assumption in it’s Meta-CoT but the final output provided does not include those steps. We believe this might be one type of artifact that arises from training with final answer verification only.
ALT Nebius AI Cloud Unveiled
