{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2025:Q7UB5XFR7FQCLRXIRFZJRIBGCW","short_pith_number":"pith:Q7UB5XFR","canonical_record":{"source":{"id":"2512.07407","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-12-08T10:39:38Z","cross_cats_sorted":[],"title_canon_sha256":"8d5fb81c3dc72e44c8c988bdb8d7ede54bf789cea04382f0e3c10503335c8ca8","abstract_canon_sha256":"89a3e8783444e4545f3861edc16aa0a7516999393d1b0d9d54a7b29f9f9a1d73"},"schema_version":"1.0"},"canonical_sha256":"87e81edcb1f96025c6e8897298a02615be91c17c1fec92947ff2bcd5f8514e68","source":{"kind":"arxiv","id":"2512.07407","version":3},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2512.07407","created_at":"2026-06-26T01:15:47Z"},{"alias_kind":"arxiv_version","alias_value":"2512.07407v3","created_at":"2026-06-26T01:15:47Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2512.07407","created_at":"2026-06-26T01:15:47Z"},{"alias_kind":"pith_short_12","alias_value":"Q7UB5XFR7FQC","created_at":"2026-06-26T01:15:47Z"},{"alias_kind":"pith_short_16","alias_value":"Q7UB5XFR7FQCLRXI","created_at":"2026-06-26T01:15:47Z"},{"alias_kind":"pith_short_8","alias_value":"Q7UB5XFR","created_at":"2026-06-26T01:15:47Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2025:Q7UB5XFR7FQCLRXIRFZJRIBGCW","target":"record","payload":{"canonical_record":{"source":{"id":"2512.07407","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-12-08T10:39:38Z","cross_cats_sorted":[],"title_canon_sha256":"8d5fb81c3dc72e44c8c988bdb8d7ede54bf789cea04382f0e3c10503335c8ca8","abstract_canon_sha256":"89a3e8783444e4545f3861edc16aa0a7516999393d1b0d9d54a7b29f9f9a1d73"},"schema_version":"1.0"},"canonical_sha256":"87e81edcb1f96025c6e8897298a02615be91c17c1fec92947ff2bcd5f8514e68","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-26T01:15:47.846421Z","signature_b64":"Dh3PiQHqttZglvRydFf4iJ1J9tQnUk05iIU5nKJHh0cKFMbYKPUdUV5RR0gVPpOPTt1mkVPwBZQgtP7mmp+RCw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"87e81edcb1f96025c6e8897298a02615be91c17c1fec92947ff2bcd5f8514e68","last_reissued_at":"2026-06-26T01:15:47.845834Z","signature_status":"signed_v1","first_computed_at":"2026-06-26T01:15:47.845834Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2512.07407","source_version":3,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-06-26T01:15:47Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"HqjEqvHW5bLIsA47VPU7O9pJgorOA6aRA/nO2Fkbl3jew2sJ5Dtc4Y6y3mKJhqtAcaRyhwu7uA52S/cnjyiPAg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-29T21:55:59.438229Z"},"content_sha256":"079193e52cf15a50dd63d6d7f82e966826e125818f23f54a5d83abd748ef6c2c","schema_version":"1.0","event_id":"sha256:079193e52cf15a50dd63d6d7f82e966826e125818f23f54a5d83abd748ef6c2c"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2025:Q7UB5XFR7FQCLRXIRFZJRIBGCW","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Training Language Models to Use Prolog as a Tool","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Training language models to use Prolog as a tool uncovers a trade-off where reward focus on correctness yields higher accuracy but delegates reasoning to natural language, while symbolic rewards enforce auditable full programs at lower peak","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Lukas Galke Poech, Niklas Mellgren, Peter Schneider-Kamp","submitted_at":"2025-12-08T10:39:38Z","abstract_excerpt":"Language models frequently produce plausible yet incorrect reasoning traces that are difficult to verify. We investigate fine-tuning models to use Prolog as an external symbolic reasoning tool, training Qwen2.5-3B-Instruct with Group Relative Policy Optimization (GRPO) on a cleaned version of GSM8K (which we release as gsm8k-prolog-prover). We systematically vary prompt structure, reward composition (execution, syntax, semantics, structure), and inference protocol (single-try, multiple-try, and two agentic modes). Our reinforcement learning approach outperforms supervised fine-tuning on GSM8K,"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We identify an accuracy--auditability trade-off: configurations tuned for correctness alone learn to delegate reasoning to natural language and use Prolog only for the final computation, while configurations rewarded for symbolic structure produce fully auditable programs at a cost in accuracy.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the observed behavioral difference between reward compositions is caused primarily by the reward signals rather than by model size limits, prompt engineering details, or the specific Prolog execution environment.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Fine-tuning Qwen2.5-3B with GRPO on GSM8K to use Prolog yields competitive zero-shot MMLU performance but exposes an accuracy-auditability trade-off interpreted as reward hacking.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Training language models to use Prolog as a tool uncovers a trade-off where reward focus on correctness yields higher accuracy but delegates reasoning to natural language, while symbolic rewards enforce auditable full programs at lower peak","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"fb36e882a2fbe7ccac754e9125f0d8d33830f336cbb9d82d9a0280a459ad9831"},"source":{"id":"2512.07407","kind":"arxiv","version":3},"verdict":{"id":"57ccc59c-d253-4c1a-b8bf-c4c603180828","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T01:12:23.389449Z","strongest_claim":"We identify an accuracy--auditability trade-off: configurations tuned for correctness alone learn to delegate reasoning to natural language and use Prolog only for the final computation, while configurations rewarded for symbolic structure produce fully auditable programs at a cost in accuracy.","one_line_summary":"Fine-tuning Qwen2.5-3B with GRPO on GSM8K to use Prolog yields competitive zero-shot MMLU performance but exposes an accuracy-auditability trade-off interpreted as reward hacking.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the observed behavioral difference between reward compositions is caused primarily by the reward signals rather than by model size limits, prompt engineering details, or the specific Prolog execution environment.","pith_extraction_headline":"Training language models to use Prolog as a tool uncovers a trade-off where reward focus on correctness yields higher accuracy but delegates reasoning to natural language, while symbolic rewards enforce auditable full programs at lower peak"},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2512.07407/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":1,"snapshot_sha256":"20b73c3d60fe461d5921f109116ade6b841c97df80cf5838b5468d989130890c"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"57ccc59c-d253-4c1a-b8bf-c4c603180828"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-06-26T01:15:47Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"b74zPeOu9PkB91EHGon3KKzN1btvEvV60jVZoODSrypsfOkwI7NjviWqmuny0HE+ZBcoUUgApC9hr1VZPq5kCQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-29T21:55:59.438701Z"},"content_sha256":"e4afe8426ea83f63545c64779d26630604c1fac8d3ea78a58717523afd0a7a52","schema_version":"1.0","event_id":"sha256:e4afe8426ea83f63545c64779d26630604c1fac8d3ea78a58717523afd0a7a52"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/Q7UB5XFR7FQCLRXIRFZJRIBGCW/bundle.json","state_url":"https://pith.science/pith/Q7UB5XFR7FQCLRXIRFZJRIBGCW/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/Q7UB5XFR7FQCLRXIRFZJRIBGCW/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":"2026-06-29T21:55:59Z","links":{"resolver":"https://pith.science/pith/Q7UB5XFR7FQCLRXIRFZJRIBGCW","bundle":"https://pith.science/pith/Q7UB5XFR7FQCLRXIRFZJRIBGCW/bundle.json","state":"https://pith.science/pith/Q7UB5XFR7FQCLRXIRFZJRIBGCW/state.json","well_known_bundle":"https://pith.science/.well-known/pith/Q7UB5XFR7FQCLRXIRFZJRIBGCW/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2025:Q7UB5XFR7FQCLRXIRFZJRIBGCW","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":"89a3e8783444e4545f3861edc16aa0a7516999393d1b0d9d54a7b29f9f9a1d73","cross_cats_sorted":[],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-12-08T10:39:38Z","title_canon_sha256":"8d5fb81c3dc72e44c8c988bdb8d7ede54bf789cea04382f0e3c10503335c8ca8"},"schema_version":"1.0","source":{"id":"2512.07407","kind":"arxiv","version":3}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2512.07407","created_at":"2026-06-26T01:15:47Z"},{"alias_kind":"arxiv_version","alias_value":"2512.07407v3","created_at":"2026-06-26T01:15:47Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2512.07407","created_at":"2026-06-26T01:15:47Z"},{"alias_kind":"pith_short_12","alias_value":"Q7UB5XFR7FQC","created_at":"2026-06-26T01:15:47Z"},{"alias_kind":"pith_short_16","alias_value":"Q7UB5XFR7FQCLRXI","created_at":"2026-06-26T01:15:47Z"},{"alias_kind":"pith_short_8","alias_value":"Q7UB5XFR","created_at":"2026-06-26T01:15:47Z"}],"graph_snapshots":[{"event_id":"sha256:e4afe8426ea83f63545c64779d26630604c1fac8d3ea78a58717523afd0a7a52","target":"graph","created_at":"2026-06-26T01:15:47Z","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":"We identify an accuracy--auditability trade-off: configurations tuned for correctness alone learn to delegate reasoning to natural language and use Prolog only for the final computation, while configurations rewarded for symbolic structure produce fully auditable programs at a cost in accuracy."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the observed behavioral difference between reward compositions is caused primarily by the reward signals rather than by model size limits, prompt engineering details, or the specific Prolog execution environment."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Fine-tuning Qwen2.5-3B with GRPO on GSM8K to use Prolog yields competitive zero-shot MMLU performance but exposes an accuracy-auditability trade-off interpreted as reward hacking."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Training language models to use Prolog as a tool uncovers a trade-off where reward focus on correctness yields higher accuracy but delegates reasoning to natural language, while symbolic rewards enforce auditable full programs at lower peak"}],"snapshot_sha256":"fb36e882a2fbe7ccac754e9125f0d8d33830f336cbb9d82d9a0280a459ad9831"},"formal_canon":{"evidence_count":1,"snapshot_sha256":"20b73c3d60fe461d5921f109116ade6b841c97df80cf5838b5468d989130890c"},"integrity":{"available":true,"clean":true,"detectors_run":[],"endpoint":"/pith/2512.07407/integrity.json","findings":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"Language models frequently produce plausible yet incorrect reasoning traces that are difficult to verify. We investigate fine-tuning models to use Prolog as an external symbolic reasoning tool, training Qwen2.5-3B-Instruct with Group Relative Policy Optimization (GRPO) on a cleaned version of GSM8K (which we release as gsm8k-prolog-prover). We systematically vary prompt structure, reward composition (execution, syntax, semantics, structure), and inference protocol (single-try, multiple-try, and two agentic modes). Our reinforcement learning approach outperforms supervised fine-tuning on GSM8K,","authors_text":"Lukas Galke Poech, Niklas Mellgren, Peter Schneider-Kamp","cross_cats":[],"headline":"Training language models to use Prolog as a tool uncovers a trade-off where reward focus on correctness yields higher accuracy but delegates reasoning to natural language, while symbolic rewards enforce auditable full programs at lower peak","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-12-08T10:39:38Z","title":"Training Language Models to Use Prolog as a Tool"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2512.07407","kind":"arxiv","version":3},"verdict":{"created_at":"2026-05-17T01:12:23.389449Z","id":"57ccc59c-d253-4c1a-b8bf-c4c603180828","model_set":{"reader":"grok-4.3"},"one_line_summary":"Fine-tuning Qwen2.5-3B with GRPO on GSM8K to use Prolog yields competitive zero-shot MMLU performance but exposes an accuracy-auditability trade-off interpreted as reward hacking.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Training language models to use Prolog as a tool uncovers a trade-off where reward focus on correctness yields higher accuracy but delegates reasoning to natural language, while symbolic rewards enforce auditable full programs at lower peak","strongest_claim":"We identify an accuracy--auditability trade-off: configurations tuned for correctness alone learn to delegate reasoning to natural language and use Prolog only for the final computation, while configurations rewarded for symbolic structure produce fully auditable programs at a cost in accuracy.","weakest_assumption":"That the observed behavioral difference between reward compositions is caused primarily by the reward signals rather than by model size limits, prompt engineering details, or the specific Prolog execution environment."}},"verdict_id":"57ccc59c-d253-4c1a-b8bf-c4c603180828"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:079193e52cf15a50dd63d6d7f82e966826e125818f23f54a5d83abd748ef6c2c","target":"record","created_at":"2026-06-26T01:15:47Z","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":"89a3e8783444e4545f3861edc16aa0a7516999393d1b0d9d54a7b29f9f9a1d73","cross_cats_sorted":[],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-12-08T10:39:38Z","title_canon_sha256":"8d5fb81c3dc72e44c8c988bdb8d7ede54bf789cea04382f0e3c10503335c8ca8"},"schema_version":"1.0","source":{"id":"2512.07407","kind":"arxiv","version":3}},"canonical_sha256":"87e81edcb1f96025c6e8897298a02615be91c17c1fec92947ff2bcd5f8514e68","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"87e81edcb1f96025c6e8897298a02615be91c17c1fec92947ff2bcd5f8514e68","first_computed_at":"2026-06-26T01:15:47.845834Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-06-26T01:15:47.845834Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"Dh3PiQHqttZglvRydFf4iJ1J9tQnUk05iIU5nKJHh0cKFMbYKPUdUV5RR0gVPpOPTt1mkVPwBZQgtP7mmp+RCw==","signature_status":"signed_v1","signed_at":"2026-06-26T01:15:47.846421Z","signed_message":"canonical_sha256_bytes"},"source_id":"2512.07407","source_kind":"arxiv","source_version":3}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:079193e52cf15a50dd63d6d7f82e966826e125818f23f54a5d83abd748ef6c2c","sha256:e4afe8426ea83f63545c64779d26630604c1fac8d3ea78a58717523afd0a7a52"],"state_sha256":"fc3d02ca2e9182c08b076fc003e1aa1b61c7ca00f2c96db317b205f926f88435"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"pogU0KfEBQW6UJfarojcjWXirbiYAyczBYLPMX4lWyEE5POShZ/1fpJ9zEUuODOJLpIBpmChQs5+r5QZne1kDw==","signed_message":"bundle_sha256_bytes","signed_at":"2026-06-29T21:55:59.440880Z","bundle_sha256":"aa8652f8d1529d64ce773e70276a50d885bc78a661cc92e02fa6895026211c76"}}