ICLR Blogpost Track 2025
Factual context validation
Organizes retrieved context into concise, verifiable facts to improve answer validation and reduce storage demands.
Ryquo Research
Questions worth
pursuing.
Advancing how AI reasons, earns trust, and works in the real world.
Featured work.
Ideas in practice
Cleanlab Blog 2025
Correcting errors in AI agents
Scores each agent message for trustworthiness and uses targeted revision or human escalation to reduce failures on Tau²-Bench.
NeurIPS Workshop Math-AI 2024
Generative sampling for robust routing
Trains neural routing solvers on generatively sampled problems to improve robustness to unfamiliar distributions, especially difficult cases.
Selected publications.
8 publications
2026
Compositional Consistency-Guided Decoding for Three-Way Logical Question Answering (opens in a new tab)
Read paperAbout this work: Compositional Consistency-Guided Decoding for Three-Way Logical Question Answering
Checks a statement against its negation to improve logical consistency and distinguish missing evidence from unnecessary abstention.
2026
Answer Only as Precisely as Justified: Calibrated Claim-Level Specificity Control for Agentic Systems (opens in a new tab)
Read paperAbout this work: Answer Only as Precisely as Justified: Calibrated Claim-Level Specificity Control for Agentic Systems
Adjusts each claim's specificity to the evidence, preserving useful detail without committing to unsupported precision.
2026
Permutation-Consensus Listwise Judging for Robust Factuality Evaluation (opens in a new tab)
Read paperAbout this work: Permutation-Consensus Listwise Judging for Robust Factuality Evaluation
Combines factuality judgments across reordered answer sets to reduce position bias and select more reliable answers.
2026
CounterRefine: Answer-Conditioned Counterevidence Retrieval for Inference-Time Knowledge Repair in Factual Question Answering (opens in a new tab)
Read paperAbout this work: CounterRefine: Answer-Conditioned Counterevidence Retrieval for Inference-Time Knowledge Repair in Factual Question Answering
Retrieves evidence that could challenge an initial answer, then revises it only when that evidence supports a correction.
2026
Beyond the Answer: Decoding the Behavior of LLMs as Scientific Reasoners (opens in a new tab)
Read paperAbout this work: Beyond the Answer: Decoding the Behavior of LLMs as Scientific Reasoners
Uses prompt optimization to reveal the reasoning patterns behind scientific problem solving and their limits across models.
2026
PAVE: Premise-Aware Validation and Editing for Retrieval-Augmented LLMs (opens in a new tab)
Read paperAbout this work: PAVE: Premise-Aware Validation and Editing for Retrieval-Augmented LLMs
Extracts explicit facts from retrieved context and uses them to check and revise answers before they are finalized.
2025
Structured Reasoning for Fairness: A Multi-Agent Approach to Bias Detection in Textual Data (opens in a new tab)
Read paperAbout this work: Structured Reasoning for Fairness: A Multi-Agent Approach to Bias Detection in Textual Data
Separates fact from opinion before assessing textual bias, pairing each judgment with a concise explanation.
2025
Trustworthy LLM Document Processing with Unstructured and Cleanlab (opens in a new tab)
Read articleAbout this work: Trustworthy LLM Document Processing with Unstructured and Cleanlab
Combines structured PDF extraction with answer-level trust scoring to make document-based AI responses easier to assess.