{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:7GCNP5QP72FF3U54YXHRRZORJ7","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"9076f40e44ed1a7dcce7ae6da57701989e6497de5d209beca364a609cc3e529b","cross_cats_sorted":["cs.AI"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LO","submitted_at":"2026-02-05T08:35:01Z","title_canon_sha256":"75016f1e0a7d0c1baf16a53aac6d508421057089eae60d44d8b5c67c7dd4701e"},"schema_version":"1.0","source":{"id":"2602.11202","kind":"arxiv","version":3}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2602.11202","created_at":"2026-05-18T02:45:05Z"},{"alias_kind":"arxiv_version","alias_value":"2602.11202v3","created_at":"2026-05-18T02:45:05Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2602.11202","created_at":"2026-05-18T02:45:05Z"},{"alias_kind":"pith_short_12","alias_value":"7GCNP5QP72FF","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"7GCNP5QP72FF3U54","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"7GCNP5QP","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:9b2950746be4b815b53f8e4662c630b556cbaeaae4a49e4ff51fa4ab18ac94ce","target":"graph","created_at":"2026-05-18T02:45:05Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"On reasoning benchmarks where policies encode mathematical or logical constraints, interwhen achieves near-perfect accuracy for reasoning models using a fraction of the tokens of baselines. On agentic benchmarks with policy-based verifier generation, it enables improvements in task quality for SLMs without any finetuning, e.g., task completion rate of Qwen3-30B jumps from 32% to 87% on the telecom domain in tau2-bench."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the monitoring system can reliably poll and fork inference to recover accurate intermediate states from any reasoning trace, and that automatic synthesis from natural-language policies produces verifiers that are both correct and sufficiently complete to catch relevant violations."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"interwhen is a single-trajectory test-time verification system that polls reasoning traces, forks inference for intermediate states, synthesizes verifiers from policies including in Lean and z3, and steers models to near-perfect accuracy and higher task completion on benchmarks."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Interwhen monitors reasoning traces in real time and steers models by verifying intermediate states against synthesized policy rules."}],"snapshot_sha256":"c6615ab1160bba7b2b6714b4c1abd43cf05e0f0d2668ba5d02f5cf232b710976"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"92441b70d55915ae497f04ec9152a0af55205b9baaa80730a175e34594828e1d"},"paper":{"abstract_excerpt":"Reasoning models produce long traces of intermediate decisions and tool calls, making test-time verification important for ensuring correctness. Existing approaches either verify only the final answer, which misses early errors, or rely on branch-and-verify strategies that explore multiple trajectories. We introduce interwhen, a single-trajectory verification framework that steers model behavior by providing feedback on intermediate reasoning traces. It addresses two key challenges. First, given a set of verifiers, obtaining verifiable states from the reasoning trace typically requires prompt ","authors_text":"Amit Sharma, Ashmit Khandelwal, Maitreyi Swaroop, Nagarajan Natarajan, Prateek Chanda, Subbarao Kambhampati, Vijval Ekbote, Vineeth N. Balasubramanian, Vishak K Bhat","cross_cats":["cs.AI"],"headline":"Interwhen monitors reasoning traces in real time and steers models by verifying intermediate states against synthesized policy rules.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LO","submitted_at":"2026-02-05T08:35:01Z","title":"interwhen: A Generalizable Framework for Steering Reasoning Models with Test-time Verification"},"references":{"count":25,"internal_anchors":0,"resolved_work":25,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"Early stopping chain-of-thoughts in large language models.ArXiv, abs/2509.14004","work_id":"ec8d501d-19f8-40d2-967e-127c68240742","year":2025},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"- A set of features (e.g., color, name, pet, book genre)","work_id":"e28c4709-c0e2-4236-b4e6-526d3c22911c","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":6,"title":"Use any feedback to guide your reasoning until a complete solution is reached","work_id":"b132ada5-6526-4e66-9fc8-58349569a657","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":7,"title":"Do not stop responding until you’ve assigned each and every variable. # Final Answer Reporting Format ‘‘‘json { \"House 1\": { \"feature1\": \"value1\", \"feature2\": \"value2\", ... }, \"House 2\": { \"feature1\":","work_id":"911a9f3a-6ba2-449c-8bce-c5d6ab69ba59","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":8,"title":"- A set of features (e.g., color, name, pet, book genre)","work_id":"75c56592-5c82-423a-8508-c55f06bec873","year":null}],"snapshot_sha256":"84a961bf08c26a5fd666b95fcb463559557691579308004e16e93d9fb29df750"},"source":{"id":"2602.11202","kind":"arxiv","version":3},"verdict":{"created_at":"2026-05-16T07:24:50.951421Z","id":"29831cd7-5e9e-4728-b65e-53960f13bf65","model_set":{"reader":"grok-4.3"},"one_line_summary":"interwhen is a single-trajectory test-time verification system that polls reasoning traces, forks inference for intermediate states, synthesizes verifiers from policies including in Lean and z3, and steers models to near-perfect accuracy and higher task completion on benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Interwhen monitors reasoning traces in real time and steers models by verifying intermediate states against synthesized policy rules.","strongest_claim":"On reasoning benchmarks where policies encode mathematical or logical constraints, interwhen achieves near-perfect accuracy for reasoning models using a fraction of the tokens of baselines. On agentic benchmarks with policy-based verifier generation, it enables improvements in task quality for SLMs without any finetuning, e.g., task completion rate of Qwen3-30B jumps from 32% to 87% on the telecom domain in tau2-bench.","weakest_assumption":"That the monitoring system can reliably poll and fork inference to recover accurate intermediate states from any reasoning trace, and that automatic synthesis from natural-language policies produces verifiers that are both correct and sufficiently complete to catch relevant violations."}},"verdict_id":"29831cd7-5e9e-4728-b65e-53960f13bf65"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:dd00c14991cad529c1e9d959013bf1d214b83c5e08a5b51040845e4affbb3bb3","target":"record","created_at":"2026-05-18T02:45:05Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"9076f40e44ed1a7dcce7ae6da57701989e6497de5d209beca364a609cc3e529b","cross_cats_sorted":["cs.AI"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LO","submitted_at":"2026-02-05T08:35:01Z","title_canon_sha256":"75016f1e0a7d0c1baf16a53aac6d508421057089eae60d44d8b5c67c7dd4701e"},"schema_version":"1.0","source":{"id":"2602.11202","kind":"arxiv","version":3}},"canonical_sha256":"f984d7f60ffe8a5dd3bcc5cf18e5d14ff0433282228ef2e479f7d5b5314e194e","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"f984d7f60ffe8a5dd3bcc5cf18e5d14ff0433282228ef2e479f7d5b5314e194e","first_computed_at":"2026-05-18T02:45:05.313031Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-18T02:45:05.313031Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"0lWGh3uXDVxdGxXi38xYAXaSMNu0sUJ4phHf/q8NWSHNM606x474dZ7jSfn4JLdy34M2iIrVXr1vbREIp00HBQ==","signature_status":"signed_v1","signed_at":"2026-05-18T02:45:05.313526Z","signed_message":"canonical_sha256_bytes"},"source_id":"2602.11202","source_kind":"arxiv","source_version":3}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:dd00c14991cad529c1e9d959013bf1d214b83c5e08a5b51040845e4affbb3bb3","sha256:9b2950746be4b815b53f8e4662c630b556cbaeaae4a49e4ff51fa4ab18ac94ce"],"state_sha256":"5a3472ea4ab533ee3e6f01f21fe5293cf2342ac7f2ec357b261abb57e1d321a9"}