{"paper":{"title":"X-SYNTH: Beyond Retrieval -- Enterprise Context Synthesis from Observed Human Attention","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Enterprise context synthesis succeeds by deriving relevance from human attention traces instead of retrieving stored system state.","cross_cats":["cs.IR","cs.LG"],"primary_cat":"cs.AI","authors_text":"George Nychis, Guruprasad Raghavan, Rohan Narayana Murthy","submitted_at":"2026-05-15T00:54:02Z","abstract_excerpt":"In enterprise operations, the context required for an AI agent task is scattered across systems of record, static information stores, and communication channels. What is stored is system state, a lossy representation of the work that actually happened [2, 52]. The prevailing approach [17, 31, 34, 36] retrieves by matching request content to what is stored; for narrow requests this works well. But synthesis quality depends on knowing what to surface and how to interpret it: knowledge specific to each organization, team, and individual [5, 57, 61], present in behavioral patterns, absent from any"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Enterprise context synthesis is not a retrieval problem. It is a relevance problem, and human attention is its most reliable ground truth. On a sales lead identification task, a frontier model unaided achieves 9.5% True Lead Rate (TLR) with 90.5% False Lead Rate (FLR). Augmented with X-SYNTH, TLR rises to 61.9% (6.5x) while FLR falls to 18.8%.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"Behavioral traces preceding positive outcomes are distinguishable from those that did not, without external labeling, allowing implicit reward signals in the data to identify causally relevant activity signatures.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"X-SYNTH synthesizes enterprise context from human behavioral attention traces modeled as Digital Twin Signatures using seven per-individual attention filters, raising true lead rate from 9.5% to 61.9% on a sales identification task.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Enterprise context synthesis succeeds by deriving relevance from human attention traces instead of retrieving stored system state.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"1397cecf8ebc026abe7d01b04f9fcc2665aa8e6d07aa33a3ab5b5b93d8db5edb"},"source":{"id":"2605.15505","kind":"arxiv","version":1},"verdict":{"id":"1af2ce9e-52ce-47e5-80c8-fbd74ddbe1b1","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T15:53:45.435925Z","strongest_claim":"Enterprise context synthesis is not a retrieval problem. It is a relevance problem, and human attention is its most reliable ground truth. On a sales lead identification task, a frontier model unaided achieves 9.5% True Lead Rate (TLR) with 90.5% False Lead Rate (FLR). Augmented with X-SYNTH, TLR rises to 61.9% (6.5x) while FLR falls to 18.8%.","one_line_summary":"X-SYNTH synthesizes enterprise context from human behavioral attention traces modeled as Digital Twin Signatures using seven per-individual attention filters, raising true lead rate from 9.5% to 61.9% on a sales identification task.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"Behavioral traces preceding positive outcomes are distinguishable from those that did not, without external labeling, allowing implicit reward signals in the data to identify causally relevant activity signatures.","pith_extraction_headline":"Enterprise context synthesis succeeds by deriving relevance from human attention traces instead of retrieving stored system state."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.15505/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-19T16:04:35.997623Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T16:01:17.938079Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"cited_work_retraction","ran_at":"2026-05-19T14:51:55.575300Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T14:21:54.060147Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"shingle_duplication","ran_at":"2026-05-19T13:49:41.852007Z","status":"skipped","version":"0.1.0","findings_count":0},{"name":"citation_quote_validity","ran_at":"2026-05-19T13:49:41.390899Z","status":"skipped","version":"0.1.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T13:33:22.638767Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"3089752f18d3f2d3a90d4b31e5923666348dda7ad69ebdeac66af34886736ced"},"references":{"count":76,"sample":[{"doi":"","year":2006,"title":"Eugene Agichtein, Eric Brill, and Susan Dumais. 2006. Improving Web Search Ranking by Incorporating User Behavior Information. InACM SIGIR Conference on Research and Development in Information Retriev","work_id":"cee6010d-2bb1-4d6c-9a74-fc4ac33c64b5","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Anonymous. 2025. ENTROPHY: Multi-Modal User Interaction Data from Live Enterprise Business Workflows. InAdvances in Neural Information Processing Systems (NeurIPS). Workfabric AI / Soroco","work_id":"a0120ba7-7a43-494a-a660-27a74c19495b","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2024. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. InInternational Conference on Learning Re","work_id":"fa6de1ad-98e4-4361-84dc-ce5255a64d9b","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2005,"title":"Jakob E. Bardram. 2005. Activity-Based Computing: Support for Mobility and Collaboration in Ubiquitous Computing.Personal and Ubiquitous Computing9, 5 (2005), 312–322","work_id":"10fa04b5-6e18-4f48-8af2-3ab4df8633d8","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2012,"title":"Paul N. Bennett, Ryen W. White, Wei Chu, Susan T. Dumais, Peter Bailey, Fedor Borisyuk, and Xiaoyuan Cui. 2012. Modeling the Impact of Short- and Long-Term Behavior on Search Personalization. InACM SI","work_id":"16af3dc4-1544-4a7e-8a07-1176e6c26b25","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":76,"snapshot_sha256":"22acc01b2c31c75b3e50386d7821e41616ec2fc4b5581020bd549608466191b4","internal_anchors":2},"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"}