Drawoffer models user career states using evidence linked reasoning
DrawOffer: Evidence-Grounded Longitudinal User-State Modeling for Career Support
Artificial Intelligence
Summary
Career-support systems need to understand when users describe their skills and experiences versus when there is actual proof of those abilities. The authors present DrawOffer, a system that separates extracting this information from text and reasoning over trusted evidence, maintaining uncertain estimates that can change with new information. Their tests using fictional personas show that DrawOffer can make accurate judgments and identify issues based on evidence, improving over simpler approaches. The work highlights the importance of grounding advice systems in verifiable evidence instead of only language model confidence.
What this means in practice
- •For career coaching platforms: Improve skill and experience evaluations of users by linking claims to verifiable evidence over time for better career advice.$Commercial implications: Enables more reliable AI-driven career support products that build trust through evidence-based assessments.
- •For human resources software teams: Incorporate evidence-grounded user profiles to support ongoing employee ability tracking and development decisions.
Authors
Lele Xu, Yuning Bian, Yifan Yuan
Abstract
Career-support assistants must distinguish user claims from evidence of current ability and withhold judgments when evidence is insufficient. DrawOffer separates language-model observation extraction from deterministic inference over source-linked records to maintain uncertain ability estimates, experience assessments, and revisable issues. We evaluate 102 fictional personas with 580 scored checkpoints, a historical full-context and Mem0 comparison, and three auxiliary knowledge-tracing datasets. On 37 test personas, DrawOffer passes 169/210 checkpoints; checkpoint-defined issue precision and recall are 0.875 and 0.636. Across three development simulations, 87.0% of checkpoint verdicts agree, including persistent failures. A cache-only development ablation replacing answer-evidence requirements with confidence-only settlement reduces majority-vote passes from 306 to 293 of 370 and issue precision from 0.927 to 0.725. Equalizing positive source weights improves predictive metrics without changing total passes; removing decay gives mixed results. These findings support an exploratory systems contribution centered on evidence sufficiency. Developer-authored scenarios, unmatched baseline models, historical test exposure, and incomplete public artifacts limit generalization and independent reproduction.