{"type":"post","slug":"full-bandwidth-transformers","title":"Full-bandwidth Transformers","description":"A preprint widens the narrow vertical channel between decoding steps by feeding the previous top-layer hidden state back in, letting non-verbalized computation re-enter the stack.","retrieval_nugget":"A preprint widens the narrow vertical channel between decoding steps by feeding the previous top-layer hidden state back in, letting non-verbalized computation re-enter the stack.","published_at":"2026-09-01","updated_at":"2026-09-12","record_date":"2026-08-25","date_kind":"discovered_at","topics":["agent-security","agents","ai"],"entities":["arXiv"],"source_url":"https://arxiv.org/abs/2608.08888","source_title":"Full-bandwidth transformer","source_domain":"arxiv.org","source_terms":["Full-bandwidth","Transformers","feedback","transformer","latent"],"summary_word_count":183,"schema_version":"newruntime-agent-readable-v0.2","stable_id":"post:full-bandwidth-transformers","status":"published","source_urls":["https://arxiv.org/abs/2608.08888"],"visuals":[],"editorial_provenance":{"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:8c465a1ca74cb36a643c8e486bdb0a5bc78fba374fc530b7fc9b2c4577127215","reviewed_at":"2026-09-12T10:00:00Z","source_evidence_count":1,"verified_claim_count":2,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":2,"implication_count":1,"watch_condition_count":1,"related_record_count":2},"analysis":{"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"A preprint submitted on 9 August points at a structural asymmetry in autoregressive transformers.","observed_facts":[{"text":"The paper describes the vertical feedback channel between decoding steps as narrow because only the sampled token returns to the bottom of the stack.","source_urls":["https://arxiv.org/abs/2608.08888"]},{"text":"Latent feedback fuses the previous top-layer hidden state with the sampled token embedding through a gated linear unit.","source_urls":["https://arxiv.org/abs/2608.08888"]}],"mechanism":"at each decoding step the previous top-layer hidden state is fused with the sampled token embedding through a gated linear unit and fed back as the next input","why_now":"Only the sampled token returns to the bottom of the stack; the top-layer hidden state is discarded.","implications":["In other words, whatever the model worked out but did not put into a token stops being thrown away at each step."],"evidence_boundary":"This is an abstract: the mechanism is described without the evaluation results, scale, or training cost that would establish whether the wider channel pays for itself.","watch_conditions":["The condition to watch is whether latent feedback survives at frontier scale under a published ablation, because a decoding-time change that helps small models often disappears when depth is already abundant."],"related_records":[{"url":"https://newruntime.com/posts/liquid-ai-releases-cpu-friendly-long-context-encoders","relation":"Prior coverage of carrying more usable state per unit of compute."},{"url":"https://newruntime.com/topics/ai","relation":"Topic hub for model architecture results."}]},"routes":{"html":"https://newruntime.com/posts/full-bandwidth-transformers/","markdown":"https://newruntime.com/posts/full-bandwidth-transformers.md","json":"https://newruntime.com/posts/full-bandwidth-transformers.json"},"editor":{"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]},"next_reads":[{"type":"topic","path":"/topics/agent-security/","url":"https://newruntime.com/topics/agent-security/","title":"Agent Security","reason":"Explore the agent-security topic hub.","media_type":"text/html","markdown_url":"https://newruntime.com/topics/agent-security.md","json_url":"https://newruntime.com/topics/agent-security.json"},{"type":"topic","path":"/topics/agents/","url":"https://newruntime.com/topics/agents/","title":"Agents","reason":"Explore the agents topic hub.","media_type":"text/html","markdown_url":"https://newruntime.com/topics/agents.md","json_url":"https://newruntime.com/topics/agents.json"},{"type":"topic","path":"/topics/ai/","url":"https://newruntime.com/topics/ai/","title":"Ai","reason":"Explore the ai topic hub.","media_type":"text/html","markdown_url":"https://newruntime.com/topics/ai.md","json_url":"https://newruntime.com/topics/ai.json"},{"type":"collection","path":"/posts/","url":"https://newruntime.com/posts/","title":"Field Notes","reason":"Open the latest editorial analysis.","media_type":"text/html","json_url":"https://newruntime.com/posts.json"},{"type":"collection","path":"/source-ledger.json","url":"https://newruntime.com/source-ledger.json","title":"Source Ledger","reason":"Inspect the public source evidence graph.","media_type":"application/json"}]}