{"paper":{"title":"Does Engram Do Memory Retrieval in Autoregressive Image Generation?","license":"http://creativecommons.org/licenses/by/4.0/","headline":"The Engram module in autoregressive image generation acts as a gated side-pathway rather than a content-addressed memory retriever.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Chunbin Gu, Jinghao Wang, Pheng-Ann Heng, Qiyuan He","submitted_at":"2026-05-13T08:40:46Z","abstract_excerpt":"The Engram module -- a hash-keyed, O(1) associative memory injected into Transformer layers -- was recently shown to improve large language model pretraining, with the appealing interpretation that it provides a content-addressed shortcut to recurring local token patterns. We ask whether this interpretation transfers to autoregressive (AR) image generation, or whether the observed gains, if any, come from a different mechanism. We adapt the Engram module to vision with 2D spatial $n$-gram hashing, gated fusion, and KV-cache-compatible incremental inference, and inject it into a class-condition"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"these findings indicate that the Engram in AR image generation behaves not as a content-addressed retriever but as a gated architectural side-pathway: a hash-keyed residual stream whose benefit is dominated by the pathway itself, with the learned table contributing only a small distributional refinement.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The 2D spatial n-gram hashing and gated fusion adaptations faithfully preserve the original Engram mechanism while allowing fair comparison to the pure AR baseline; if these changes fundamentally alter how the module operates, the mechanistic conclusions would not hold.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Engram in AR image generation saves backbone FLOPs but trails pure AR baselines in FID and behaves as a gated side-pathway rather than a content-addressed retriever.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"The Engram module in autoregressive image generation acts as a gated side-pathway rather than a content-addressed memory retriever.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"ff39da7fea8e52fea58fc5852193896870475734db8b8cc6738898dda385a8d6"},"source":{"id":"2605.13179","kind":"arxiv","version":1},"verdict":{"id":"a365a755-4d3b-4811-8a83-6e8a9cd168c5","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:30:04.590990Z","strongest_claim":"these findings indicate that the Engram in AR image generation behaves not as a content-addressed retriever but as a gated architectural side-pathway: a hash-keyed residual stream whose benefit is dominated by the pathway itself, with the learned table contributing only a small distributional refinement.","one_line_summary":"Engram in AR image generation saves backbone FLOPs but trails pure AR baselines in FID and behaves as a gated side-pathway rather than a content-addressed retriever.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The 2D spatial n-gram hashing and gated fusion adaptations faithfully preserve the original Engram mechanism while allowing fair comparison to the pure AR baseline; if these changes fundamentally alter how the module operates, the mechanistic conclusions would not hold.","pith_extraction_headline":"The Engram module in autoregressive image generation acts as a gated side-pathway rather than a content-addressed memory retriever."},"references":{"count":19,"sample":[{"doi":"","year":2022,"title":"Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. InCVPR, 2022","work_id":"021352e7-6e83-4d54-9e3f-99bdf6ccabc2","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models","work_id":"942d7453-d7f7-4ad9-8180-e04bd226f6e5","ref_index":2,"cited_arxiv_id":"2601.07372","is_internal_anchor":true},{"doi":"","year":2009,"title":"Imagenet: A large-scale hierarchical image database","work_id":"a54af14a-31f5-43fe-bd58-cc2bbda964d7","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Taming transformers for high-resolution image synthesis","work_id":"0f792ff9-97c6-4511-9b2d-b30263cfe7f1","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2014,"title":"Neural Turing Machines","work_id":"0a5ce53c-9670-42b9-8be5-386de7eed50c","ref_index":5,"cited_arxiv_id":"1410.5401","is_internal_anchor":true}],"resolved_work":19,"snapshot_sha256":"dbeee697967ff6b6284a778dc78a0e545b109b7d7060b70b94900d010ab37b2a","internal_anchors":4},"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"}