01
Evaluating relationships between facts
BeyondAtoms examines outputs whose individual facts hold but whose relationships are distorted. Its benchmark covers relation injection, omission, re-anchoring and substitution. CARVE uses structure extraction and bidirectional relation alignment to detect errors missed by isolated fact checks.
BeyondAtoms · EMNLP 2026QuestionCorrect individual facts do not guarantee a correct account.
MethodEvaluate relationships as a distinct object of verification.
02
Making factuality verification reusable
E-Verify separates decomposition, embedding and interaction, moving fine-grained verification into embedding space. Subsequent Gated Residual Discrepancy research uses dual-perspective residuals and adaptive gating to isolate factual differences while reusing precomputed evidence representations.
E-Verify · Findings of EMNLP 2025QuestionJoint encoding repeatedly processes the same evidence.
MethodSeparate evidence representation from fine-grained interaction.
03
Turning expert judgment into executable workflows
Vibe Calibration organizes measurement commands, parameter constraints, acceptance criteria and audit records into decision-tree Skills. Agents choose subsequent measurements and recover from anomalies while retaining raw data, fits and decisions for inspection.
Vibe Calibration · arXiv 2026QuestionScripts automate operations, but decisions often remain tacit.
MethodPackage decision criteria, tool interfaces and recovery paths together.
04
Bringing scientific constraints into search
CrystalFormer-CSP studies crystal structure prediction conditioned on chemical composition, combining a generative model with subsequent search. Team experience also includes materials-model adaptation, PPO/DPO post-training and workflows for candidate analysis and experiment scheduling.
CrystalFormer-CSP · arXiv 2025QuestionLarge candidate spaces must be constrained by scientific requirements.
MethodCombine generation and search with computational feedback.
This includes work undertaken by the team in prior research collaborations. EMNLP 2026 acceptance information is based on team-provided records.
