{ "schema_version": 1, "builder_plan": { "scenario": "You already need task-specific training and want to reduce its memory or compute cost.", "change": "Change only the training technique on a small controlled run after checking prerequisites.", "baseline": "Your current training recipe with the same model, data split, and evaluation.", "steps": [ "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.", "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.", "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.", "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test." ], "measure": "Held-out quality, peak memory, training time, and full compute cost to the target quality.", "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation." }, "builder_review": null, "arxiv_id": "2609.26355", "title": "PACT: From Credit Assignment to Critic Alignment", "summary": "PACT replaces the standard simultaneous actor-critic update in PPO-style RL post-training with an Actor-then-Critic order plus importance-sampling correction on critic training, so the value function is aligned with the already-updated policy. The cost is moderate: an extra ordered update step and correction logic within an existing actor-critic pipeline, but it still requires RL fine-tuning infrastructure, a trainable critic, and reward or teacher signals, which strains a single 24 GB GPU for larger models. It can fail if the importance-sampling correction introduces gradient variance, if reward sparsity assumptions break outside math and code tasks, or if the gains do not transfer from the reported agentic benchmarks to the reader's domain.", "rationale": "The method is an incremental modification to PPO-style training with plausible quality gains, but it presupposes RL post-training capacity that only fits constrained infrastructure for small models or via API-based rollouts, so it warrants a scoped pilot rather than immediate adoption.", "technique": "Actor-then-Critic PPO variant with aligned critic (PACT)", "family": "treino_eficiente", "recommendation": "testar", "published": "2026-09-22", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.26355", "independent_implementations": 0, "total_implementations": 1, "stars": 20, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "PACT replaces the standard simultaneous actor-critic update in PPO-style RL post-training with an Actor-then-Critic order plus importance-sampling correction on critic training, so the value function is aligned with the already-updated policy. The cost is moderate: an extra ordered update step and correction logic within an existing actor-critic pipeline, but it still requires RL fine-tuning infrastructure, a trainable critic, and reward or teacher signals, which strains a single 24 GB GPU for larger models. It can fail if the importance-sampling correction introduces gradient variance, if reward sparsity assumptions break outside math and code tasks, or if the gains do not transfer from the reported agentic benchmarks to the reader's domain.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.26355", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Reports +8.80 accuracy points over GRPO and +13.16 over PPO averaged across four math benchmarks, and +2.0 to +3.8 pass-rate points over PPO, GRPO, and SAO on SWE-bench Verified.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.26355", "source_page": null, "source_excerpt": "" } ], "exposure_map": [ { "dimension": "quality", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compute", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "latency", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "operations", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compatibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "security", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "data_and_training", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "reproducibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" } ], "risks": [], "minimum_test": [], "open_questions": [], "independent_tests": [], "lab_reproductions": [], "report_available": false, "validation_tier": "unknown", "evidence_tier": "unknown", "infrastructure_basis": "unknown", "equations": [], "equations_status": "not_fetched", "core_kind": "", "equations_fetched_at": "", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }