I work on AI agents, reinforcement learning, agentic memory, and document intelligence, combining research with building practical systems. My recent work includes RDAN-GRPO for instruction following, notSoSmartOCR for extracting structured content from complex documents, and LevelHead for training-free page deskewing and orientation classification. I work at HiLabs and am associated with the Vision and Learning Group at IIT Roorkee.
I graduated from IIT Roorkee in 2026, previously worked on generative engine optimisation at Info Edge, and contributed to GNU Octave through Google Summer of Code. Outside work, I run, swim, and powerlift.
My papers Mixing Matters, on position bias across Transformer, Mamba, and hybrid language models, and Post-Audit Mirage, on the identifiability limits of offline agent verification, were accepted at the PriGM and Who Verifies the Agents? workshops at NeurIPS 2026. MycoGate-VL, on multilingual hallucination detection, was accepted at SHROOM-Visions at EMNLP 2026.
For more details, reach out at aman.behera.systems@gmail.com.
Updated
Built a document-intelligence product that turns complex scans into structured, reviewable outputs. It combines orientation correction, layout detection, Falcon OCR 1.5, and selective LLM refinement for handwriting and ambiguous regions, and shows page-level evidence, per-block confidence, tables, and Markdown output.
To be submitted · CVPR 2027
A training-free method for page deskewing and orientation classification. LevelHead levels the page with a spectral deskewer first, then lets any frozen classifier or VLM vote over the four quarter turns of the level page. On pages at any angle it lifts four Qwen3-VL models from 41.9% to 84.6% and four orientation classifiers from 71.9% to 85.2%.
Accepted · SHROOM-Visions at EMNLP 2026
Detects and categorizes hallucinated character spans in vision-language responses across English, French, Italian, and Chinese. Combines multilingual detectors, a learned clean-document gate, visual category supervision, and scorer-aware output serialization.
Accepted · PriGM Workshop at NeurIPS 2026, Paris
A paired study of position bias across Transformer, Mamba, and hybrid language models. One answer-bearing passage moves through ten positions of a fixed long context while everything else stays the same. Pythia favours the opening and the Mamba checkpoints do not; the paper reports this as a model-family difference, not an isolated architecture effect.
Accepted · Who Verifies the Agents? Workshop at NeurIPS 2026 (Poster)
A matched-world benchmark for offline agent verification. A safe and a harmful deployment can leave byte-identical audit records while needing opposite decisions, so no offline verifier can tell them apart. Across authorization, scheduling, and shared-capacity environments, the paper shows when fresh post-update evidence restores the decision, and gives a fail-closed monitor that deploys, holds, asks for more evidence, or reports unsupported.
Rubric Decoupled Advantage Normalisation GRPO is a rubric-based reinforcement learning repository for aligning LLMs with complex, open-domain instruction-following tasks. It improves performance on public instruction-following benchmarks.
Developed an interview practice platform that generates diverse, topic-specific questions and uses NLP to analyse spoken responses, identify grammatical errors, and provide corrections after the interview.
Identifiability limits of offline agent verification, with Payal Payal. Poster. Paper and code.
Training-free page deskewing and orientation classification for any model, to be submitted to CVPR 2027. Project page and code.
Accepted at the Workshop on Principles of Generative Modeling. A paired study of position bias across Transformer, Mamba, and hybrid language models. Project page, paper, and code.
Multilingual hallucination detection in vision-language responses. Paper and code.
Document extraction with layout detection, OCR, selective LLM refinement, and visual evidence for review. Explore the project.
Contributed to the open-source scientific computing framework.
Led a team of six to a first runner-up finish on the ISRO geospatial VLM problem statement among 24 IITs.
Received after winning the HiLabs AIQuest hackathon at IIT Roorkee.
Worked on generative engine optimisation (GEO) at Info Edge.
Built an AI interview platform.