{"record": {"aliases": ["Evaluating LLM Uncertainty in Long-Form Generation Using Deterministic Ground Truth"], "artifacts": [{"id": "arxiv:2607.03870", "kind": "paper", "url": "https://arxiv.org/abs/2607.03870"}, {"id": "gh:idoamit198/salt", "kind": "repo", "url": "https://github.com/IdoAmit198/SALT"}], "authors": ["Ido Amit", "Ido Galil", "Ran El-Yaniv"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.03870", "source_url": "https://arxiv.org/abs/2607.03870"}, "categories": ["General AI", "Language & Knowledge", "cs.AI"], "description": {"en": "SALT is a procedurally generated benchmark for evaluating correctness and uncertainty estimates in long-form language-model outputs against deterministic ground truth."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.03870", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.03870", "title": "Single-answer Atomic Long-form Target (SALT)"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2607.03870", "modality": null, "name": "Single-answer Atomic Long-form Target (SALT)", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-09-25T04:40:22Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_evaluating-llm-uncertainty-in-long-form-ge_b5aeae69", "origin_source": "arxiv", "recognition_confidence": "0.7", "record_sha256": "dce8775a98deb807495868101dcbe53ad36a73586d301316c7898db987602d6a", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T17:23:20.811378Z", "source_url": "https://arxiv.org/abs/2607.03870"}, "publisher": null, "released": "2026-07-04", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.03870", "source_url": "https://arxiv.org/abs/2607.03870"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-03870", "source": "claire_radar", "source_benchmark_id": "2607.03870", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Evaluating LLM Uncertainty in Long-Form Generation Using Deterministic Ground Truth"], "annotation": "Machine Generated", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "Deterministic long-form uncertainty evaluation and an ICML 2026 acceptance provide a strong basis for near-term technical attention.", "score": 4}, "capabilities": [], "confidence": "Medium", "construction": "Original Synthetic", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T17:23:20.811378Z", "inputHash": "3ebe760f8e597aae3eb8fb4305bc950b96a3d92a383eaeae1a8dd7a120795ae2", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We introduce Single-answer Atomic Long-form Target (SALT), a benchmark", "canonicalNameSource": "abstract", "decisionReason": "The supplied evidence provides a formal benchmark name, repeatable deterministic tasks, explicit evaluation dimensions, and a public code path.", "evaluationEvidence": {"reuse": "Code available at https://github.com/IdoAmit198/SALT", "scoring": "unit-level evaluation of correctness, calibration, and ranking without external judges", "task": "six procedurally generated tasks with single deterministic long textual ground truths"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T17:23:20.811378Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "SALT is a procedurally generated benchmark for evaluating correctness and uncertainty estimates in long-form language-model outputs against deterministic ground truth.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We introduce Single-answer Atomic Long-form Target (SALT), a benchmark of six procedurally generated tasks with single deterministic long textual ground truths, enabling unit-level evaluation of correctness, calibration, and ranking without external judges."}, "familyId": "bmf_e427cb3df73b", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_evaluating-llm-uncertainty-in-long-form-ge_b5aeae69", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/IdoAmit198/SALT", "data": null, "paper": "https://arxiv.org/abs/2607.03870", "pdf": "https://arxiv.org/pdf/2607.03870", "project": null, "report": "https://arxiv.org/abs/2607.03870"}, "metrics": [], "motivation": "As LLMs generate increasingly long outputs, effective uncertainty estimation must identify errors at fine-grained levels rather than discard entire responses.", "name": "Single-answer Atomic Long-form Target (SALT)", "oneLine": "SALT is a procedurally generated benchmark for evaluating correctness and uncertainty estimates in long-form language-model outputs against deterministic ground truth.", "paperTitle": "Evaluating LLM Uncertainty in Long-Form Generation Using Deterministic Ground Truth", "primaryDomain": "General AI", "publication": {"evidence": "Accepted to the 43rd International Conference on Machine Learning (ICML 2026). Code available at https://github.com/IdoAmit198/SALT", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2607.03870", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "43rd International Conference on Machine Learning (ICML 2026)", "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Runnable", "recognitionConfidence": 0.7, "relation": "introduces", "releasedAt": "2026-07-04", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Ido Amit", "Ido Galil", "Ran El-Yaniv"], "categories": ["cs.AI", "cs.CL", "cs.LG"], "id": "2607.03870", "title": "Evaluating LLM Uncertainty in Long-Form Generation Using Deterministic Ground Truth", "type": "arxiv", "url": "https://arxiv.org/abs/2607.03870"}, "sourceUpdatedAt": "2026-07-04T13:25:18Z", "topics": ["cs.AI"], "venueAttempts": [{"decisionRaw": "Accepted to the 43rd International Conference on Machine Learning (ICML 2026). Code available at https://github.com/IdoAmit198/SALT", "evidence": [{"level": "author-claim", "observedAt": "2026-08-19T11:05:13.395391Z", "rawValue": "Accepted to the 43rd International Conference on Machine Learning (ICML 2026). Code available at https://github.com/IdoAmit198/SALT", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2607.03870"}], "reviewStatus": "accepted", "venueName": "43rd International Conference on Machine Learning (ICML 2026)"}], "whyItMatters": "It permits fine-grained, judge-free comparison of accuracy, calibration, and confidence ranking across output resolutions."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
