ZAFARYAB HAIDER

Topology-aware world models for electricity-grid reconfiguration — ranking line-switching candidates via learned imagined rollouts, verified by exact AC power-flow simulation.

Whether token-level logits, probabilities and generation-time signals from autoregressive LLMs give reliable evidence of incorrect or unsupported outputs.

Semantically targeted sparse parameter bit flips in speech-recognition models, with robustness and failure analysis.

Anchor-calibrated trusted/corrupted cohort recovery when corrupted sources form the majority.

Local-to-system harm amplification in multi-agent LLM systems — how compromised agents propagate failures.

Sparse bit-level fault analysis and semantic steering in quantized vision-language models.

Consensus-based reward mitigating malicious human-in-the-loop feedback during RLHF training of LLMs.

Meta-attack framework composing existing attack primitives to synthesize effective and stealthy adversarial examples.