Knowledge‑intensive language‑model systems usually encode external knowledge as text chunks or static graphs, which limits support for concept evolution, point‑in‑time reasoning, and the distinction between validated and inferred knowledge. To address this, we introduce the Concept Lifecycle Model (CLM), which treats concepts as persistent, graph‑grounded, temporally versioned entities with explicit provenance and epistemic status. Building on CLM, we propose Concept‑Grounded Attention (CGA), injecting the concept graph into Transformer computation in two forms:
- Form A: graph‑biased self‑attention that steers attention according to graph structure;
- Form B: gated cross‑attention over concept nodes that fuses concept information.
We evaluate the framework in controlled settings using disabled‑mechanism baselines. On 200 MuSiQue and HotpotQA questions with retrieval fixed, concept‑graph retrieval recovers explicit multi‑hop paths but does not improve evidence recall. Form A directs 2.76× more attention to gold concepts than distractors, yet the same ratio appears when Form A is disabled, indicating a negligible learned bias and no answer changes. An identity‑preserving Form B raises F1 from 0.188 to 0.221, but control concepts achieve 0.213, suggesting most of the gain stems from added capacity.
On LongMemEval, explicit temporal representation improves answer accuracy by 13–25 points across all tested generators up to 122B parameters. A simplified CLM version resolution performs similarly to dated serialization because concept identity is not reliably established. In a synthetic source‑independence task, protocol‑derived epistemic status reduces unsupported assertions from 28% to 0.1% in a fine‑tuned small model and from 19–68% to 0–5% in 72–122B models.
Overall, the results support making temporal validity and epistemic status explicit, while showing that graph‑attention diagnostics are not informative without disabled‑mechanism controls.
Review: The CLM offers a dual perspective on knowledge—temporal and provenance—while CGA’s two injection mechanisms exhibit distinct contributions. Form B’s capacity effect warrants deeper investigation.