{"paper":{"title":"Debunking Grad-ECLIP: A Comprehensive Study on Its Incorrectness and Fundamental Principles for Model Interpretation","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Grad-ECLIP produces model interpretations that do not match the original model's behavior or performance because its method is equivalent to a simpler attention-based route.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Xiaohui Fan, Yongjin Cui","submitted_at":"2026-05-13T03:35:23Z","abstract_excerpt":"Grad-ECLIP is published at ICML 2024 and represents a new Transformer interpretation technical route (intermediate features-based). First, this paper demonstrates that the intermediate features-based technical route is not a novel one. Based on the existing attention-based route, we have developed Attention-ECLIP, which is completely equivalent to Grad-ECLIP but with simpler computation. Both through formal derivation and experimental validation, we prove that the intermediate feature-based route represented by Grad-ECLIP is actually an equivalent variant of the attention-based route. Next, th"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Both through formal derivation and experimental validation, we prove that the intermediate feature-based route represented by Grad-ECLIP is actually an equivalent variant of the attention-based route. ... the model interpretation results obtained by Grad-ECLIP are not those of the original model, and the interpretation results are misaligned with the model's performance.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the authors' Attention-ECLIP is exactly equivalent in all practical cases and that their experiments correctly isolate misalignment without selection bias or implementation differences.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Grad-ECLIP is an equivalent but flawed variant of attention-based interpretation, with two principles proposed to ensure model explanations reflect the original model.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Grad-ECLIP produces model interpretations that do not match the original model's behavior or performance because its method is equivalent to a simpler attention-based route.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"3bf2bd6f626c800427260f2afab06e52fc83222ed6e393afec09b08e68c14f17"},"source":{"id":"2605.12952","kind":"arxiv","version":1},"verdict":{"id":"69bd1621-e0ca-4fd8-bbe6-ce773b4fbec2","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:49:53.312777Z","strongest_claim":"Both through formal derivation and experimental validation, we prove that the intermediate feature-based route represented by Grad-ECLIP is actually an equivalent variant of the attention-based route. ... the model interpretation results obtained by Grad-ECLIP are not those of the original model, and the interpretation results are misaligned with the model's performance.","one_line_summary":"Grad-ECLIP is an equivalent but flawed variant of attention-based interpretation, with two principles proposed to ensure model explanations reflect the original model.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the authors' Attention-ECLIP is exactly equivalent in all practical cases and that their experiments correctly isolate misalignment without selection bias or implementation differences.","pith_extraction_headline":"Grad-ECLIP produces model interpretations that do not match the original model's behavior or performance because its method is equivalent to a simpler attention-based route."},"references":{"count":43,"sample":[{"doi":"10.18653/v1/2020.acl-main.385","year":2020,"title":"Quantifying attention flow in transformers","work_id":"f25a0031-4d07-43be-a809-e5e972089225","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1145/3583780.3614836","year":2023,"title":"Deep integrated explanations","work_id":"4f7c5182-6d18-4b6a-96b3-c4637396a8fe","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"Interpretability via model extrac- tion","work_id":"3b0433b4-6917-4328-a479-3f4ddc6169e2","ref_index":3,"cited_arxiv_id":"1706.09773","is_internal_anchor":true},{"doi":"10.1007/978-3-319-44781-0","year":2016,"title":"Layer-wise relevance propagation for neural networks with local renormalization layers","work_id":"63381708-2650-4f7a-a38d-6a8a4710a432","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1109/cvpr46437.2021.00356","year":2021,"title":"Adabins: Depth estimation using adap- tive bins","work_id":"7083a41e-5666-435b-ab26-c753f6490b9a","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":43,"snapshot_sha256":"aaca938599b03026b75b396fa187eb1491ff3913b00cfa9a3d72e81a245b2663","internal_anchors":6},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}