{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:EYGETUQ36HEDM3ITQT66ERFYH5","short_pith_number":"pith:EYGETUQ3","canonical_record":{"source":{"id":"2605.03195","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-04T22:24:24Z","cross_cats_sorted":["cs.SE"],"title_canon_sha256":"56f945d5cd08749c7fd37a09ea400d0fdaedbf6db6dddc849eb9664f373dd95c","abstract_canon_sha256":"a160c8cbab67bd12133a831288435df182caf357b5934e2c8b64a41a4d856371"},"schema_version":"1.0"},"canonical_sha256":"260c49d21bf1c8366d1384fde244b83f60daf6bc5e7c31f4090949de478dc2c6","source":{"kind":"arxiv","id":"2605.03195","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.03195","created_at":"2026-07-09T00:19:14Z"},{"alias_kind":"arxiv_version","alias_value":"2605.03195v2","created_at":"2026-07-09T00:19:14Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.03195","created_at":"2026-07-09T00:19:14Z"},{"alias_kind":"pith_short_12","alias_value":"EYGETUQ36HED","created_at":"2026-07-09T00:19:14Z"},{"alias_kind":"pith_short_16","alias_value":"EYGETUQ36HEDM3IT","created_at":"2026-07-09T00:19:14Z"},{"alias_kind":"pith_short_8","alias_value":"EYGETUQ3","created_at":"2026-07-09T00:19:14Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:EYGETUQ36HEDM3ITQT66ERFYH5","target":"record","payload":{"canonical_record":{"source":{"id":"2605.03195","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-04T22:24:24Z","cross_cats_sorted":["cs.SE"],"title_canon_sha256":"56f945d5cd08749c7fd37a09ea400d0fdaedbf6db6dddc849eb9664f373dd95c","abstract_canon_sha256":"a160c8cbab67bd12133a831288435df182caf357b5934e2c8b64a41a4d856371"},"schema_version":"1.0"},"canonical_sha256":"260c49d21bf1c8366d1384fde244b83f60daf6bc5e7c31f4090949de478dc2c6","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-09T00:19:14.002012Z","signature_b64":"7C+K3JyF/LzwVAXi9+Ogu552tRV+IBzIwWxs+N0YPTo+eeSH1Ty+fKxmVNA1zIdUJ4+txjpME/pOlR/uEK4PCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"260c49d21bf1c8366d1384fde244b83f60daf6bc5e7c31f4090949de478dc2c6","last_reissued_at":"2026-07-09T00:19:14.001205Z","signature_status":"signed_v1","first_computed_at":"2026-07-09T00:19:14.001205Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2605.03195","source_version":2,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-07-09T00:19:14Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"7kANO5eY2obBRMIhrdHhSX0CThoAkX4ETr2wSUy7Ov+VKeXUDnkSVqQb/bjcz9IvoXCGOiXIz5g4SNbWGM37CQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-04T04:28:15.843385Z"},"content_sha256":"d7bcd5153ca9f24a5d568847141625d5cfd80288892540bb207d37e38f6b1e36","schema_version":"1.0","event_id":"sha256:d7bcd5153ca9f24a5d568847141625d5cfd80288892540bb207d37e38f6b1e36"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:EYGETUQ36HEDM3ITQT66ERFYH5","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Terminus-4B: Can a Smaller Model Replace Frontier LLMs at Agentic Execution Tasks?","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"A 4B model can match frontier LLMs at agentic terminal execution tasks","cross_cats":["cs.SE"],"primary_cat":"cs.AI","authors_text":"Spandan Garg, Vikram Nitin, Yufan Huang","submitted_at":"2026-05-04T22:24:24Z","abstract_excerpt":"Modern coding agents increasingly delegate specialized subtasks to subagents, which are smaller, focused agentic loops that handle narrow responsibilities like search, debugging or terminal execution. This architectural pattern keeps the main agent's context window clean by isolating verbose outputs (e.g. build logs, test results, etc.) within the subagent context. Typically when agents employ subagents for such tasks, they use frontier models as these subagents. In this paper, we investigate whether a finetuned small language model (SLM) can achieve comparable performance to frontier models i"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Terminus-4B is able to reduce the token usage of the main agent by up to ~30% compared to the No Subagent baseline with no impact to agent performance on benchmarks like SWE-Bench Pro and our internal SWE-Bench C# benchmark... Furthermore, Terminus-4B improves key metrics showing the main agent relying on the outputs of the subagent and doing fewer terminal execution tasks by itself. We see that our model not only closes the gap between the Vanilla Qwen model and frontier models like Claude Sonnet / Opus / GPT-5.3-Codex, but often even exceeds their performance.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the rubric-based LLM-as-judge reward produces training signals that generalize beyond the specific benchmarks and do not embed biases from the judge model itself.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A fine-tuned 4B model matches or exceeds frontier LLMs in terminal execution subagent tasks for coding agents, reducing main agent token usage by 30% with no performance loss.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A 4B model can match frontier LLMs at agentic terminal execution tasks","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"3937de43f443fcecaa32bbe2b691ab5a0e72b26a38d7c0d2bc7b2fafda658d1a"},"source":{"id":"2605.03195","kind":"arxiv","version":2},"verdict":{"id":"3628de00-9059-48fa-a4c3-31232b97e40c","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T18:05:19.549400Z","strongest_claim":"Terminus-4B is able to reduce the token usage of the main agent by up to ~30% compared to the No Subagent baseline with no impact to agent performance on benchmarks like SWE-Bench Pro and our internal SWE-Bench C# benchmark... Furthermore, Terminus-4B improves key metrics showing the main agent relying on the outputs of the subagent and doing fewer terminal execution tasks by itself. We see that our model not only closes the gap between the Vanilla Qwen model and frontier models like Claude Sonnet / Opus / GPT-5.3-Codex, but often even exceeds their performance.","one_line_summary":"A fine-tuned 4B model matches or exceeds frontier LLMs in terminal execution subagent tasks for coding agents, reducing main agent token usage by 30% with no performance loss.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the rubric-based LLM-as-judge reward produces training signals that generalize beyond the specific benchmarks and do not embed biases from the judge model itself.","pith_extraction_headline":"A 4B model can match frontier LLMs at agentic terminal execution tasks"},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.03195/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T14:35:41.810177Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-20T01:31:21.958406Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T15:35:11.900686Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"18410850d72f0c749519697b6e4bb622fd19bc769983a594b0a6f2a97a70eb57"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"33261308317bab0af441fb6222dba8815f5039d5355d2749b132ef677bfc838a"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"3628de00-9059-48fa-a4c3-31232b97e40c"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-07-09T00:19:14Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"PiB9MMVO2x3IfBqbt7vsiDFK0Y4BgKpiu3JOypHVae8qZHyLxQdPKj31hWYNbL7UxkQSntnlnQIuLDQp47UXAQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-08-04T04:28:15.844083Z"},"content_sha256":"4f96ccdb9e99b251452bef8df3dbef234291b3e2c4e7bd1fe33911b68d493ed9","schema_version":"1.0","event_id":"sha256:4f96ccdb9e99b251452bef8df3dbef234291b3e2c4e7bd1fe33911b68d493ed9"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/EYGETUQ36HEDM3ITQT66ERFYH5/bundle.json","state_url":"https://pith.science/pith/EYGETUQ36HEDM3ITQT66ERFYH5/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/EYGETUQ36HEDM3ITQT66ERFYH5/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-08-04T04:28:15Z","links":{"resolver":"https://pith.science/pith/EYGETUQ36HEDM3ITQT66ERFYH5","bundle":"https://pith.science/pith/EYGETUQ36HEDM3ITQT66ERFYH5/bundle.json","state":"https://pith.science/pith/EYGETUQ36HEDM3ITQT66ERFYH5/state.json","well_known_bundle":"https://pith.science/.well-known/pith/EYGETUQ36HEDM3ITQT66ERFYH5/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:EYGETUQ36HEDM3ITQT66ERFYH5","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":"a160c8cbab67bd12133a831288435df182caf357b5934e2c8b64a41a4d856371","cross_cats_sorted":["cs.SE"],"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-04T22:24:24Z","title_canon_sha256":"56f945d5cd08749c7fd37a09ea400d0fdaedbf6db6dddc849eb9664f373dd95c"},"schema_version":"1.0","source":{"id":"2605.03195","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.03195","created_at":"2026-07-09T00:19:14Z"},{"alias_kind":"arxiv_version","alias_value":"2605.03195v2","created_at":"2026-07-09T00:19:14Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.03195","created_at":"2026-07-09T00:19:14Z"},{"alias_kind":"pith_short_12","alias_value":"EYGETUQ36HED","created_at":"2026-07-09T00:19:14Z"},{"alias_kind":"pith_short_16","alias_value":"EYGETUQ36HEDM3IT","created_at":"2026-07-09T00:19:14Z"},{"alias_kind":"pith_short_8","alias_value":"EYGETUQ3","created_at":"2026-07-09T00:19:14Z"}],"graph_snapshots":[{"event_id":"sha256:4f96ccdb9e99b251452bef8df3dbef234291b3e2c4e7bd1fe33911b68d493ed9","target":"graph","created_at":"2026-07-09T00:19:14Z","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":"Terminus-4B is able to reduce the token usage of the main agent by up to ~30% compared to the No Subagent baseline with no impact to agent performance on benchmarks like SWE-Bench Pro and our internal SWE-Bench C# benchmark... Furthermore, Terminus-4B improves key metrics showing the main agent relying on the outputs of the subagent and doing fewer terminal execution tasks by itself. We see that our model not only closes the gap between the Vanilla Qwen model and frontier models like Claude Sonnet / Opus / GPT-5.3-Codex, but often even exceeds their performance."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the rubric-based LLM-as-judge reward produces training signals that generalize beyond the specific benchmarks and do not embed biases from the judge model itself."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"A fine-tuned 4B model matches or exceeds frontier LLMs in terminal execution subagent tasks for coding agents, reducing main agent token usage by 30% with no performance loss."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"A 4B model can match frontier LLMs at agentic terminal execution tasks"}],"snapshot_sha256":"3937de43f443fcecaa32bbe2b691ab5a0e72b26a38d7c0d2bc7b2fafda658d1a"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"33261308317bab0af441fb6222dba8815f5039d5355d2749b132ef677bfc838a"},"integrity":{"available":true,"clean":true,"detectors_run":[{"findings_count":0,"name":"ai_meta_artifact","ran_at":"2026-05-20T14:35:41.810177Z","status":"completed","version":"1.0.0"},{"findings_count":0,"name":"doi_title_agreement","ran_at":"2026-05-20T01:31:21.958406Z","status":"completed","version":"1.0.0"},{"findings_count":0,"name":"doi_compliance","ran_at":"2026-05-19T15:35:11.900686Z","status":"completed","version":"1.0.0"}],"endpoint":"/pith/2605.03195/integrity.json","findings":[],"snapshot_sha256":"18410850d72f0c749519697b6e4bb622fd19bc769983a594b0a6f2a97a70eb57","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"Modern coding agents increasingly delegate specialized subtasks to subagents, which are smaller, focused agentic loops that handle narrow responsibilities like search, debugging or terminal execution. This architectural pattern keeps the main agent's context window clean by isolating verbose outputs (e.g. build logs, test results, etc.) within the subagent context. Typically when agents employ subagents for such tasks, they use frontier models as these subagents. In this paper, we investigate whether a finetuned small language model (SLM) can achieve comparable performance to frontier models i","authors_text":"Spandan Garg, Vikram Nitin, Yufan Huang","cross_cats":["cs.SE"],"headline":"A 4B model can match frontier LLMs at agentic terminal execution tasks","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-04T22:24:24Z","title":"Terminus-4B: Can a Smaller Model Replace Frontier LLMs at Agentic Execution Tasks?"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.03195","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-08T18:05:19.549400Z","id":"3628de00-9059-48fa-a4c3-31232b97e40c","model_set":{"reader":"grok-4.3"},"one_line_summary":"A fine-tuned 4B model matches or exceeds frontier LLMs in terminal execution subagent tasks for coding agents, reducing main agent token usage by 30% with no performance loss.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"A 4B model can match frontier LLMs at agentic terminal execution tasks","strongest_claim":"Terminus-4B is able to reduce the token usage of the main agent by up to ~30% compared to the No Subagent baseline with no impact to agent performance on benchmarks like SWE-Bench Pro and our internal SWE-Bench C# benchmark... Furthermore, Terminus-4B improves key metrics showing the main agent relying on the outputs of the subagent and doing fewer terminal execution tasks by itself. We see that our model not only closes the gap between the Vanilla Qwen model and frontier models like Claude Sonnet / Opus / GPT-5.3-Codex, but often even exceeds their performance.","weakest_assumption":"That the rubric-based LLM-as-judge reward produces training signals that generalize beyond the specific benchmarks and do not embed biases from the judge model itself."}},"verdict_id":"3628de00-9059-48fa-a4c3-31232b97e40c"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:d7bcd5153ca9f24a5d568847141625d5cfd80288892540bb207d37e38f6b1e36","target":"record","created_at":"2026-07-09T00:19:14Z","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":"a160c8cbab67bd12133a831288435df182caf357b5934e2c8b64a41a4d856371","cross_cats_sorted":["cs.SE"],"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-04T22:24:24Z","title_canon_sha256":"56f945d5cd08749c7fd37a09ea400d0fdaedbf6db6dddc849eb9664f373dd95c"},"schema_version":"1.0","source":{"id":"2605.03195","kind":"arxiv","version":2}},"canonical_sha256":"260c49d21bf1c8366d1384fde244b83f60daf6bc5e7c31f4090949de478dc2c6","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"260c49d21bf1c8366d1384fde244b83f60daf6bc5e7c31f4090949de478dc2c6","first_computed_at":"2026-07-09T00:19:14.001205Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-07-09T00:19:14.001205Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"7C+K3JyF/LzwVAXi9+Ogu552tRV+IBzIwWxs+N0YPTo+eeSH1Ty+fKxmVNA1zIdUJ4+txjpME/pOlR/uEK4PCQ==","signature_status":"signed_v1","signed_at":"2026-07-09T00:19:14.002012Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.03195","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:d7bcd5153ca9f24a5d568847141625d5cfd80288892540bb207d37e38f6b1e36","sha256:4f96ccdb9e99b251452bef8df3dbef234291b3e2c4e7bd1fe33911b68d493ed9"],"state_sha256":"fd374f776e35ec4f00f24d1a8ec43388261c3ff85ffdf4791e120baf48a0e82e"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"z34CDR12kMrvI7RixlhnuTQiGfgZjTeCyOsCnoehvvWkDn9dyzJZE+VuXWxwLQKLtxMGj9HS0/xgCSExvWEHDQ==","signed_message":"bundle_sha256_bytes","signed_at":"2026-08-04T04:28:15.849344Z","bundle_sha256":"d7f9b357e12a113941814c4c10168f487e8f8762ddd5af77fe12488af55e9efe"}}