current work

Ongoing research and projects

Current Role

Applied Scientist II at Amazon (January 2026 - Present)
Working on agentic bot systems.


Under Review at NeurIPS 2026

Sparsifying Correlation Clustering: Edge Coresets, Triangle Witnesses, and Observation Lower Bounds (with Sanjeda Akter, Anuj Sharma)
arXiv:2602.13684

Exclusivity Without Marginal Loss: A Coupled-Rounding Primitive and Full-LP Obstruction for Chromatic Correlation Clustering (with Sanjeda Akter, Anuj Sharma)

Topology-Aware State Abstraction with Tangle Cores for Markov Decision Processes (with Sanjeda Akter, Anuj Sharma)

Coverage-Based Calibration for Post-Training Quantization via Weighted Maximum Coverage over Outlier Channels (with Sanjeda Akter, Anuj Sharma)

Continual Calibration: Coverage Can Collapse Before Accuracy in Lifelong LLM Fine-Tuning (with Sanjeda Akter, Anuj Sharma)

Audit-Certified Conformal Routing for Reliable Hybrid Training (with Sanjeda Akter, Anuj Sharma)

Differentiable Range-Partition Entropy for Entropy-Sensitive Geometric Algorithms (with Sanjeda Akter, Anuj Sharma)
arXiv:2509.03733

Concentration-Scaled Grassmannian Routing for Controllable Mixture-of-Experts (with Sanjeda Akter, Anuj Sharma)
arXiv:2602.17798

Auditing Near-Optimal Policies Can Be Exponentially Hard: Conditional Query Lower Bounds via Occupancy Rashomon Capacity (with Sanjeda Akter, Anuj Sharma)

Minimax Round Complexity of Distributed Offline Dynamic Programming: Discounted Locality, Bit-Radius Tradeoffs, and Gossip Baselines (solo)


Under Review at AAAI 2027

Tight Online-to-PAC Conversion under Graph-Mixing Dependence (with Sanjeda Akter, Anuj Sharma)

Margin Certificates for Voting Classifiers under Mixing Dependence (with A. B. Habib)

The Cost of Adaptivity: Matching Lower Bounds Across Learning Problems (with A. B. Habib)

Graph Dimensionality Reduction for Contextual Bandits: Pathwise Misspecification Bounds under Noisy Eigenspaces (with Joyanta J. Mondal)

Discrepancy-Rounded Fair Bandits with Static and Time-Varying Exposure Floors (with Joyanta J. Mondal)

Sub-Quadratic Bisimulation Metrics via Approximate Nearest Neighbors: A Coverage-Augmented Approximation Guarantee (with Joyanta J. Mondal)

Certified Predictive Value-of-Advice Gating for Cost-Aware Language-Model Guidance in Reinforcement Learning (with M. N. Swaqeeb, Abu Sa-Adat Mohamed Moon-Im Al Ahsan)

Auditing Reinforcement Learning Explanations: Quantifier Frontiers, Strategy Representation, and Verifiable Islands (with Abu Sa-Adat Mohamed Moon-Im Al Ahsan, M. N. Swaqeeb)

Finite Constant Frontiers and Auditable Regret Certificates for Average-Reward Reinforcement Learning (with Abu Sa-Adat Mohamed Moon-Im Al Ahsan, M. N. Swaqeeb)

Adversary-Aligned Error Propagation for Robust Reinforcement Learning under Function Approximation (with M. N. Swaqeeb, Abu Sa-Adat Mohamed Moon-Im Al Ahsan)


Planned for / Under Review at NAACL 2027 (ACL Rolling Review, August 2026 Cycle)

Dynamic Proxy-Mixing: Transferring Replay Controllers from Small to Large Models for Continual Instruction Tuning (with Fariya Afrin, Anuj Sharma)

Calibration-Preserving Pruning: Compression as a Reliability Contract (with A. B. Habib, Anuj Sharma)

Selective Risk Certification for LLM Outputs via Information-Lift Statistics: PAC-Bayes, Robustness, and Skeleton Design (with Sanjeda Akter, Anuj Sharma)
arXiv:2509.12527

Spectral-Guided Diffusion: Accelerating Inference via Static Spectral Layer Scheduling (with Abu Sa-Adat Mohamed Moon-Im Al Ahsan, Anuj Sharma)

Canonicalized Stable-List Replay for Private Federated Continual Learning over Language-Model Embeddings (with Abu Sa-Adat Mohamed Moon-Im Al Ahsan, Anuj Sharma)

Calibrated Adaptation: Bayesian Stiefel Manifold Priors for Reliable Parameter-Efficient Fine-Tuning (with Sanjeda Akter, M. M. Hossain, Anuj Sharma)
arXiv:2602.17809

Accuracy Gaps Warn but Do Not Certify: Cross-Domain Recalibration of LLM Judges (with Fariya Afrin)

The Accuracy Gap Stratifies Recalibration: When Post-Hoc Calibration of LLM Judges Helps, and When It Backfires (with Fariya Afrin)

Evaluator Ensembles Under Reward Hacking: Covariance Geometry and Finite-Search Guarantees (with Fariya Afrin)

Quality-Diversity Stress Tests for Process Reward Models: What Archive Coverage Can and Cannot Certify (with Fariya Afrin)

Task-Aware Spectral Pruning: A Mixture-of-Masks Framework for Efficient LLM Inference (with Fariya Afrin, Sanjeda Akter, Anuj Sharma)

Spectral Tail Interventions in Language Models: Reasoning-Sensitive Weight Structure from Controlled Surgery (with Sanjeda Akter, M. N. Swaqeeb, Abu Sa-Adat Mohamed Moon-Im Al Ahsan, Anuj Sharma)

Learning to Forget Attention: Memory Consolidation for Adaptive Compute Reduction (with Sanjeda Akter, Anuj Sharma)
arXiv:2602.12204

CODS: Iterative Bellman-Residual Data Selection for Reusable Offline Reinforcement Learning (with Sanjeda Akter, Abu Sa-Adat Mohamed Moon-Im Al Ahsan, M. N. Swaqeeb, Anuj Sharma)

BLADE: Distilled LLM Regularization for Calibrated Knowledge Graph Completion (with R. B. Tamanna, A. El Karaky, Sanjeda Akter, Anuj Sharma)

Anytime-Valid Skeleton-Relative Lift Certificates for LLM Generation Stopping (with Sanjeda Akter, Anuj Sharma)
arXiv:2510.06478

Stateful CARS: Exact Adaptive Sampling for Policy-Constrained LLM Agents (with M. N. Swaqeeb, Abu Sa-Adat Mohamed Moon-Im Al Ahsan)

Private Anytime Selective-Risk Certification for Federated Retrieval-Augmented Generation: Guarantees and Empirical Limits (with Sanjeda Akter, Anuj Sharma)

FlowNeg: GFlowNet-Guided Diverse Hard Negative Sampling for Knowledge Graph Embedding (with N. A. Hridi, Joyanta J. Mondal)

Opportunity Is Not Realizability: Selection-Valid Diagnostics for Multi-LLM Routing (with Abu Sa-Adat Mohamed Moon-Im Al Ahsan, M. N. Swaqeeb)

StepCOPS: Closed-Testing Lower-Tail Certificates for Language-Model Policy Selection (with Sanjeda Akter, Anuj Sharma)

When Is Benchmark Contamination Detectable? Information Limits and Power-Calibrated Audits (with Sanjeda Akter, Anuj Sharma)


Under Review at IEEE T-ITS

Calibrated and Resource-Aware Super-Resolution for Reliable Driver Behavior Analysis (with Sanjeda Akter, Anuj Sharma)
arXiv:2509.23535


Preprints

Large Language Models for Crash Detection in Video: A Survey of Methods, Datasets, and Challenges (with Sanjeda Akter, Anuj Sharma)
arXiv:2507.02074


Upcoming Presentations

  • None