Mithun Paul Saint-Germain

Research profile

Research summary

Mithun Paul Saint-Germain is a researcher at Arizona State University whose 2026 output spans LLM agent security, quantum natural language processing, and LLM-guided multi-robot systems. OpenAlex indexes 4 works in the last three years, all 2026 arXiv preprints, with an h-index and i10-index of 0 — an early-stage profile resting on fresh preprints rather than an established citation record. The most coherent line is agent security: AgentDrift (last author) introduces the first public step-level benchmark of injection-hijacked LLM agent trajectories, and DriftNet (last author) builds a dual-head trajectory transformer on it, reaching F1 0.983 for detecting and localizing prompt injection, though both are self-benchmarked on synthetic, single-generator data. A robotics paper (middle author) couples one LLM per robot with UCB-guided policy refinement and local Double DQN, reporting a best median of 42 ticks in single-run simulations, while the Hindi quantum NLP study (middle author) lifts contextual backoff accuracy to 64.56% versus 11.65% for raw LLM prompting on a 206-token test set. The field distribution splits evenly at 0.25 across four areas and focus scores 0, so this reads as a broad exploratory phase; average reproducibility readiness is 57 across 4 papers, highest for the benchmark work. Collaboration is small-team: 6 distinct coauthors, with repeated partners including Asif Pinjari, Gautami Sanjay Naik, Krishna Bhatia, H Aswath Babu, Chongwen Dong, and Carlo R. daCunha.

Research keywords

Highlights

• AgentDrift: first public step-level benchmark of injection-hijacked LLM agent trajectories

  Evolve's review credits the paper with claiming the first public benchmark of step-level injection labels on LLM agent tool-call trajectories, and notes it is admirably honest about its flaws — though only a surface baseline tests the benchmark's difficulty.

• DriftNet detects and localizes prompt injection with F1 0.983

  A dual-head trajectory transformer claiming the first joint verdict-and-step-label detector for LLM agents; results are rigorous but self-benchmarked on synthetic AgentDrift data from a single generator. Reproducibility readiness: high.

• Automatic pregroup supertagging scales Hindi quantum NLP

  Contextual backoff reaches 64.56% against 11.65% for raw LLM prompting, but Evolve flags a tiny 206-token test corpus and no neural baselines. Reproducibility readiness: low.

• Schema-bounded LLM refines multi-robot policies without destabilizing local learning

  One LLM per robot with UCB-guided round-level refinement plus local Double DQN; the C4 configuration posts the best median of 42 ticks, but evidence rests on single-run, seed-free simulation. Reproducibility readiness: medium.

Timeline

1. 2026Releases a schema-bounded LLM scheme for refining multi-robot policies (arXiv 2609.05133), pairing UCB-guided round-level refinement with local Double DQN.
2. 2026Publishes AgentDrift, a step-labeled benchmark of injection-hijacked LLM agent trajectories (arXiv 2609.06972).
3. 2026Follows with DriftNet, a dual-head trajectory transformer reaching F1 0.983 for injection detection and localization (arXiv 2609.10892).
4. 2026Extends quantum NLP to Hindi via automatic pregroup supertagging, with contextual backoff at 64.56% (arXiv 2609.13721).

Papers in the Evolve catalog

• 2026-09-12

  Scaling Hindi Quantum Natural Language Processing through Automatic Pregroup Supertagging

• 2026-09-09

  DriftNet: A Dual-Head Trajectory Transformer for Detecting and Localizing Prompt Injection in LLM Agents

• 2026-09-07

  AgentDrift: A Step-Labeled Benchmark of Injection-Hijacked LLM Agent Trajectories

• 2026-09-04

  A Schema Bounded Language Model for Refining Robot Policies Without Destabilizing Local Learning

Frequent co-authors