Sham Satish Thakare — Bio

I am drawn to questions where standard empirical assumptions quietly break down when inspected carefully.

In modern post-training and reinforcement learning with verifiable rewards (RLVR), the community widely assumes that token-predictive entropy works as an effective proxy for model uncertainty. In my recent work, I set out to test whether that assumption actually holds. It turns out that predictive entropy correlates strongly with sequence length (r = +0.486)—meaning that uncertainty-weighted credit assignment methods (like entropy-weighted GRPO) end up misallocating reward gradients simply because a correct mathematical derivation required more steps. We formulated length-invariant credit assignment estimators (ear_grpo_reasoning) and state-matched difference-in-differences counterfactual evaluation protocols (recovery_eval, stateshift) to measure whether models genuinely recover from intermediate reasoning errors when structural covariates are held frozen.

On the systems side, I care just as much about runtime reality as theoretical elegance. When distributed consensus protocols encounter asymmetric network partitions or non-uniform round-trip times, conventional static quorums suffer crippling tail-latency blowups. To solve this, I designed failure-domain aware dynamic quorum adaptation algorithms and spectral latency estimators in C++20 for Raft consensus clusters (quorumshift), proving non-empty intersection invariants without ever violating safety or linearizability.

My general rule of thumb is simple: work out the math on paper first, identify the falsifiable hypothesis, and then write the low-level code and stress-test benchmarks to see if the hypothesis survives contact with reality.

Sustained research requires physical energy and a clear head. I am a competitive runner, specializing in sprints and distance running, and won the Gold Medal in Athletics at the 2023 Inter-College Athletic Meet. Most of my toughest algorithmic roadblocks get untangled while out on early morning runs across Pune's hills before the city wakes up.

During my final undergraduate year, I was elected President of the Social Volunteer Organization at APCOER. We ran weekly volunteer tutoring programs for underprivileged high-school students across Pune district, teaching them foundational algebra, arithmetic, and basic digital literacy. We also organized city-wide blood donation camps and local environmental sustainability campaigns. Sitting side-by-side with young students who had never owned a computer and watching them solve algebraic equations for the first time remains one of the most meaningful things I have done.

When I am away from terminal windows and LaTeX manuscripts, you will usually find me reading classical computer science literature, drinking strong tea, or exploring Pune's historic neighborhoods with a notebook and fountain pen.