QERA
A training-free inference-time architecture that identifies implicit requirements and retrieves evidence by posing questions to the context before answering, comprising stages such as requirement enumeration, full-dialogue retrieval and self-checking, and protected local repair.
Timeline (last 90 days)
QERA (Question-driven Evidence Reasoning Architecture) is a question-driven evidence reasoning architecture whose core idea is to question the context before answering.
QERA's workflow includes checklist-ifying requirements, full-conversation retrieval with self-checking, and protected local repair.
QERA requires no training and no predefined scoring criteria, and performs only about four LLM calls per task on average.
On all 405 tasks, QERA achieves the highest solve rate across four frontier large models, averaging 21.0%; the strongest of the three inference-time baselines is 18.2%, and direct answering is 14.6%.
QERA uses fewer tokens than PSCI, only 4% to 5% of those used by Ctx2Skill.
Ablation experiments on GPT-6-Luna show that QERA's gains are established before answer generation: the fixed answering protocol contributes 3.8 scoring points and request-specific questions contribute 1.6 scoring points.
QERA's improvements are concentrated on implicit scoring criteria, while post-hoc checking mainly refines numeric requirements.
Even without implementing the full QERA, fixing a protocol of "list requirements first, then gather evidence, then answer" for long-context question answering can yield considerable gains.
All Facts (8)
QERA (Question-driven Evidence Reasoning Architecture) is a question-driven evidence reasoning architecture whose core idea is to question the context before answering.
50%UnverifiedQERA's workflow includes checklist-ifying requirements, full-conversation retrieval with self-checking, and protected local repair.
50%UnverifiedQERA requires no training and no predefined scoring criteria, and performs only about four LLM calls per task on average.
50%UnverifiedOn all 405 tasks, QERA achieves the highest solve rate across four frontier large models, averaging 21.0%; the strongest of the three inference-time baselines is 18.2%, and direct answering is 14.6%.
50%UnverifiedQERA uses fewer tokens than PSCI, only 4% to 5% of those used by Ctx2Skill.
50%UnverifiedAblation experiments on GPT-6-Luna show that QERA's gains are established before answer generation: the fixed answering protocol contributes 3.8 scoring points and request-specific questions contribute 1.6 scoring points.
50%UnverifiedQERA's improvements are concentrated on implicit scoring criteria, while post-hoc checking mainly refines numeric requirements.
50%UnverifiedEven without implementing the full QERA, fixing a protocol of "list requirements first, then gather evidence, then answer" for long-context question answering can yield considerable gains.
50%