Keep on to blur preview images; turn off to show them clearly
RT @RazorpayTech: Excited to host @CloudflareDev this weekend in our office. Cloudflare is at the forefront of building technology at pla…
Have questions, or building something cool with Cloudflare's Developer products? We're here to help. For help with your account please try @CloudflareHelp
4/ The new version can be found at https://t.co/4LGWfiCZ5e These results will also be presented by the lead authors @karthikv792 @kayastechly & @PalodVardh12428 at #NeurIPS2025 workshops on LAW, ForLM and Efficient Reasoning next week. Please stop by and chat!
AI researcher & teacher @SCAI_ASU. Former President of @RealAAAI; Chair of @AAAS Sec T. Here to tweach #AI. YouTube Ch: https://t.co/4beUPOmf6y Bsky: rao2z
3/ Finally we include a study of the correlation between the length of intermediate tokens, and the computational complexity of the problem instance. The results show that there is no correlation between them! (I discussed an earlier version of this experiment in https://t.co/RL9ZEOKbpQ)
4/ The new version can be found at https://t.co/4LGWfiCZ5e These results will also be presented by the lead authors @karthikv792 @kayastechly & @PalodVardh12428 at #NeurIPS2025 workshops on LAW, ForLM and Efficient Reasoning next week. Please stop by and chat!
We just uploaded an expanded version of the "Beyond Semantics" paper--our systematic study of the role of intermediate tokens in LRMs--to arXiv, and it may be of interest to some of you. 🧵 1/ One intriguing new study is the effect of training the base transformer with a mix of correct and incorrect traces. We notice that as the % of incorrect (swapped) traces during training goes from 0 to 100, the trace validity of the models at inference time falls down monotonically (plot on right below) as expected, but the solution accuracy exhibits a U-curve (plot on left)! This suggests that what seems to matter is the "consistency" of the traces used during training rather than their correctness.
AI researcher & teacher @SCAI_ASU. Former President of @RealAAAI; Chair of @AAAS Sec T. Here to tweach #AI. YouTube Ch: https://t.co/4beUPOmf6y Bsky: rao2z
We just uploaded an expanded version of the "Beyond Semantics" paper--our systematic study of the role of intermediate tokens in LRMs--to arXiv, and it may be of interest to some of you. 🧵 1/ One intriguing new study is the effect of training the base transformer with a mix of correct and incorrect traces. We notice that as the % of incorrect (swapped) traces during training goes from 0 to 100, the trace validity of the models at inference time falls down monotonically (plot on right below) as expected, but the solution accuracy exhibits a U-curve (plot on left)! This suggests that what seems to matter is the "consistency" of the traces used during training rather than their correctness.
2/ We also look at the effect of DeepSeek R1-style RL on the trace validity--to see if RL improves trace validity of the base model. The results show that RL is basically neutral on trace validity. It improves solution accuracy even in the case of model trained on 100% swapped traces, without increasing trace validity.
This is the kind of cool shit we do with our sponsors that you'll help us brainstorm.
If you're interested in being a future sponsor, reach out here: https://t.co/5Bf3JvyPIx
