Research:
I work on advancing reliable reasoning, reinforcement learning credit assignment for foundation models, and failure-tolerant distributed consensus systems. My current research interests include:
- (i) Reasoning Reliability & RLVR: investigating token-predictive entropy confounding with reasoning sequence length ($r = +0.486$) in reinforcement learning with verifiable rewards (GRPO), and developing length-invariant credit assignment estimators.
- (ii) State-Matched Error Recovery: designing difference-in-differences counterfactual continuation evaluation protocols (
recovery_evalandstateshift) to measure whether language models recover from invalid intermediate derivations under frozen structural covariates across post-training checkpoints. - (iii) Distributed Consensus & Fault Tolerance: formulating failure-domain aware dynamic quorum adaptation and predictive replica selection in C++20 Raft consensus clusters (
quorumshift), eliminating tail latency spikes under asymmetric network fault injection (conducted as part of distributed consensus research).
Academic Honors & Leadership:
- First Class with Distinction (Honors): Cumulative GPA: 8.70 / 10.00; Final Year SGPA: 9.53 / 10.00, Savitribai Phule Pune University.
- President, Social Volunteer Club (2023–2024): Elected student head of university community outreach during final year; directed district-wide blood donation drives, educational mentoring for underprivileged youth, campus sustainability campaigns, and social welfare programs across Pune.
- Teaching Assistant (2023–2024): Instructional support and tutorial sessions for undergraduate coursework in Data Structures & Algorithms and Foundations of Machine Learning.
- Gold Medal, Athletics: Inter-College Athletic Meet (Sprint / Distance Running, 2023).
Recent Manuscripts & Preprints:
Selected manuscripts and preprints (also see [publications] and [GitHub]) advancing foundation model reasoning and distributed systems:
Sham Satish Thakare
Under Review at IEEE Transactions on Artificial Intelligence (IEEE TAI), 2026
Sham Satish Thakare
Working Paper, 2026 • Distributed Systems Research
Sham Satish Thakare
Under Review at 11th IEEE Special Session on Machine Learning on Big Data (IEEE BigData 2026)
[StateShift: State-Matched Evaluation of Error Recovery in Neural Reasoning]
Sham Satish Thakare
Submitted to Elsevier Artificial Intelligence (AIJ), 2026
[MediRush-SafeAgent: Layered Runtime Enforcement for Healthcare-Commerce LLM Tool Execution]
Sham Satish Thakare
In Peer Review at Intelligence-Based Medicine (Elsevier) & SSRN Preprint (ID: 7446543)
Open-Source Research Software:
[ear_grpo_reasoning] — PyTorch probing suite for token predictive entropy and sequence-length confounding controls in GRPO.
[quorumshift] — High-performance C++20 dynamic quorum adaptation for Raft consensus clusters .
[recovery_eval] — State-matched continuation evaluation protocol measuring error recovery under frozen structural covariates.
[adaptive-rl-forge] — Diagnostic probe library for foundation model checkpoint reward plasticity saving 18.7% RL training FLOPs.
[enclaveshield] — Zero-knowledge memory attestation proofs and adaptive Path ORAM tree rebalancing for hardware enclaves.
[medirush-safeagent] — Multi-stage deterministic runtime policy interception architecture for autonomous tool-using agent safety.
Recent Blog Posts:
Selected blog posts (also see [publications] and [GitHub]) exploring fundamental questions in AI and systems: