{"paper":{"title":"SinkTrack: Attention Sink based Context Anchoring for Large Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"By injecting contextual features into the attention sink token, SinkTrack keeps large language models focused on the original input throughout generation, reducing hallucination and forgetting.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Guikun Chen, Wenguan Wang, Xu Liu","submitted_at":"2026-04-11T04:49:11Z","abstract_excerpt":"Large language models (LLMs) suffer from hallucination and context forgetting. Prior studies suggest that attention drift is a primary cause of these problems, where LLMs' focus shifts towards newly generated tokens and away from the initial input context. To counteract this, we make use of a related, intrinsic characteristic of LLMs: attention sink -- the tendency to consistently allocate high attention to the very first token (i.e., <BOS>) of a sequence. Concretely, we propose an advanced context anchoring method, SinkTrack, which treats <BOS> as an information anchor and injects key context"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Experiments demonstrate that SinkTrack mitigates hallucination and context forgetting across both textual (e.g., +21.6% on SQuAD2.0 with Llama3.1-8B-Instruct) and multi-modal (e.g., +22.8% on M3CoT with Qwen2.5-VL-7B-Instruct) tasks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That consistently high attention to the BOS token can be reliably turned into an effective context anchor simply by injecting input-derived features into its representation, without side effects or the need for per-model tuning.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SinkTrack uses attention sink at the BOS token to anchor LLMs to initial context, reducing hallucination and forgetting with reported gains on benchmarks like SQuAD2.0 and M3CoT.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"By injecting contextual features into the attention sink token, SinkTrack keeps large language models focused on the original input throughout generation, reducing hallucination and forgetting.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"6e3013d3d0d18f75f2368e172817751f2fa04a37953db2bfa6d6cf945b83d036"},"source":{"id":"2604.10027","kind":"arxiv","version":2},"verdict":{"id":"6db75f4c-5a58-4212-be50-8d9f1fef30d3","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T15:43:49.317116Z","strongest_claim":"Experiments demonstrate that SinkTrack mitigates hallucination and context forgetting across both textual (e.g., +21.6% on SQuAD2.0 with Llama3.1-8B-Instruct) and multi-modal (e.g., +22.8% on M3CoT with Qwen2.5-VL-7B-Instruct) tasks.","one_line_summary":"SinkTrack uses attention sink at the BOS token to anchor LLMs to initial context, reducing hallucination and forgetting with reported gains on benchmarks like SQuAD2.0 and M3CoT.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That consistently high attention to the BOS token can be reliably turned into an effective context anchor simply by injecting input-derived features into its representation, without side effects or the need for per-model tuning.","pith_extraction_headline":"By injecting contextual features into the attention sink token, SinkTrack keeps large language models focused on the original input throughout generation, reducing hallucination and forgetting."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.10027/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"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"}