Vinay Kumar Sankarapu
Publications
ALIGNBEAM : Inference-Time Alignment Transfer via Cross-Vocabulary Logit Mixing
Domain fine-tuning degrades the safety of large language models: fine-tuned specialists readily comply with harmful prompts framed in domain language. Existing inference-time defenses that mix logits from a safe anchor model require both models to share a vocabulary, which rules them out for the cross-family specialists where safety is most degraded. We present ALIGNBEAM, a training-free method that lifts this restriction by translating anchor logits into the target model's vocabulary token-by-token at each decoding step; a small LLM judge then selects the safest among K candidate continuations. No weights are changed, and the safety-utility trade-off can be tuned at deployment without retraining. Across both cross-vocabulary and same-vocabulary evaluation pairs, ALIGNBEAM substantially raises refusal on adversarial benchmarks while keeping task accuracy and inference overhead within practical bounds. The results show that safety alignment can be transferred between model families at inference time, without touching either model's weights.
Provenance-Grounded Gating and Adaptive Recovery in Synthetic Post-Training Data Curation
Synthetic post-training pipelines commonly filter generated samples with reward models or holistic LLM judges, yet two practices remain rarely examined together: whether the filtering signal is grounded in the source evidence that induced each generation, and whether rejected samples can be systematically recovered rather than permanently discarded. We present a controlled study of both questions across gate configurations, recovery strategies, and generator scales, using adversarially injected corpora to provide ground-truth failure labels. We find that exact source provenance improves faithfulness gating for stronger judges, that hallucination and reward gates reject largely disjoint sample populations making both necessary, and that an adaptive recovery pipeline combining failure diagnosis with targeted regeneration achieves higher yield, recovery rate, and injection recall than naive resampling. Downstream fine-tuning quality is driven primarily by generator scale, with filtration and recovery conditions contributing meaningfully but secondarily.
Position: Behavioural Assurance Cannot Verify the Safety Claims Governance Now Demands
This position paper argues that behavioural assurance, even when carefully designed, is being asked to carry safety claims it cannot verify. AI governance frameworks enacted between 2019 and early 2026 require reviewable evidence of properties such as the absence of hidden objectives, resistance to loss-of-control precursors, and bounded catastrophic capability; current assurance methodologies (primarily behavioural evaluations and red-teaming) are epistemically limited to observable model outputs and cannot verify the latent representations or long-horizon agentic behaviours these frameworks presume to regulate. We formalize this structural mismatch as the audit gap, the divergence between required and achievable verification access, and introduce the concept of fragile assurance to describe cases where the evidential structure does not support the asserted safety claim. Through an analysis of a 21-instrument inventory, we identify an incentive gradient where geopolitical and industrial pressures systematically reward surface-level behavioral proxies over deep structural verification. Finally, we propose a technical pivot: bounding the weight of behavioral evidence in legal text and extending voluntary pre-deployment access with mechanistic-evidence classes, specifically linear probes, activation patching, and before/after-training comparisons.
Beyond Uniform Credit: Causal Credit Assignment for Policy Optimization
Policy gradient methods for language model reasoning, such as GRPO and DAPO, assign uniform credit to all generated tokens - the filler phrase "Let me think" receives the same gradient update as the critical calculation "23 + 45 = 68." We propose counterfactual importance weighting: mask reasoning spans, measure the drop in answer probability, and upweight tokens accordingly during policy gradient updates. Our method requires no auxiliary models or external annotation, instead importance is estimated directly from the policy model's own probability shifts. Experiments on GSM8K across three models spanning the Qwen and Llama families demonstrate consistent improvements over uniform baselines and faster convergence to equivalent accuracy. Inverting the importance signal hurts performance, confirming we capture genuine causal structure rather than noise. Analysis shows the method correctly prioritizes calculation steps over scaffolding text. We view these findings as establishing counterfactual importance weighting as a foundation for further research rather than a complete solution.
Beyond KL Divergence: Policy Optimization with Flexible Bregman Divergences for LLM Reasoning
Policy optimization methods like Group Relative Policy Optimization (GRPO) and its variants have achieved strong results on mathematical reasoning and code generation tasks. Despite extensive exploration of reward processing strategies and training dynamics, all existing group-based methods exclusively use KL divergence for policy regularization, leaving the choice of divergence function unexplored. We introduce Group-Based Mirror Policy Optimization (GBMPO), a framework that extends group-based policy optimization to flexible Bregman divergences, including hand-designed alternatives (L2 in probability space) and learned neural mirror maps. On GSM8K mathematical reasoning, hand-designed ProbL2-GRPO achieves 86.7% accuracy, improving +5.5 points over the Dr. GRPO baseline. On MBPP code generation, neural mirror maps reach 60.1-60.8% pass@1, with random initialization already capturing most of the benefit. While evolutionary strategies meta-learning provides marginal accuracy improvements, its primary value lies in variance reduction ($\pm$0.2 versus $\pm$0.6) and efficiency gains (15% shorter responses on MBPP), suggesting that random initialization of neural mirror maps is sufficient for most practical applications. These results establish divergence choice as a critical, previously unexplored design dimension in group-based policy optimization for LLM reasoning.