TRACE: Transformer Rationale Attribution via Cumulative Evidence, An In-training Mechanism for Faithful and Depth-integrated Explainability

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Transformer encoders build rich contextual representations through successive layers of self-attention, but this depth comes at a cost to explainability: a token that attracts strong attention in an early block has no structural guarantee of remaining influential in later ones, so post-hoc attribution methods cannot reliably recover which tokens the model genuinely relied upon.We introduce TRACE (Transformer Rationale Attribution via Cumulative Evidence), a parameter-free mechanism that embeds attribution directly into the encoder forward pass.At each block, TRACE extracts a scalar self-salience score for every token and maintains a decayed cumulative maximum across depth, amplifying persistently salient tokens while isolated early peaks decay.On two ERASER datasets, TRACE improves causal faithfulness over post-hoc soft-attribution baselines, reaching Comprehensiveness of 0.336 on Movie Reviews, the highest among the soft models compared, and Sufficiency of 0.047 on MultiRC, the lowest among those with interpretable scores.The gain is not uniform: token-level rationale F1 on Movie Reviews (0.2945) does not exceed an unsupervised rationalisation baseline (0.322), and classification F1 is moderately reduced by the frozen-backbone protocol.An ablation identifies depth-persistent accumulation, rather than self-salience alone, as the source of the improvement.Attribution computed during training rather than reconstructed afterwards therefore yields more faithful explanations, at a measurable cost in predictive accuracy.