​Scaling test-time computation has become a reliable lever for improving LLM performance on complex reasoning, but current approaches are wasteful. Most discard 99% of generated trajectories; others inflate token counts through compulsive "overthinking."

​This talk maps the evolution of inference-time scaling, from foundational methods like Best-of-N sampling and Self -Consistency to advanced algorithms like Tree Search, AlphaEvolve, and Recursive Self-Aggregation (RSA).

​From there, it proposes a shift: from black-box sampling to logit-driven inference control using a model's own uncertainty signals to guide search more precisely.

​Two recent frameworks illustrate this:

• ​TTSD (Test-Time Scaling via Self-Distillation Signals): Instead of global re-prompting, TTSD compares conditional log-probabilities to localise errors at the token level. Failed trajectories are truncated exactly where logic diverges, maximally reusing valid prefixes and cutting token costs by ~50%.

• ​VASP (Verified Adaptive Stopping Policies): Treats sub-argmax End-of-Thought probabilities as dynamic triggers for self-verification. Acts as an optimal stopping policy — matching majority-voting accuracy while reducing reasoning token budget by up to 50%.

​The core argument: analysing a model's internal probability distributions can turn expensive brute-force generation into efficient, state-aware reasoning.

​Speaker Bio

​Rahul Madhavan is a Researcher at Google DeepMind India, focusing on RL for LLMs. Major interest areas have been credit attribution and formalising inference time control. His recent focus has been on Open-endedness, and Recursive self improvement.

​Rahul has finished his PhD in computer science from IISc, specialising in safety in ML, RL, Causality and Optimisation.

​LinkedIn: /rahul-madhavan

​
To attend online:

• ​⁠Add to calendar: https://shorturl.at/YEsQ2

• ​Gmeet link: meet.google.com/krx-rwoe-eqv

​Looking forward to see you at event!