{ "schema": "project-experience.writing-task-review.v1", "version": "1.0", "status": "constructed_example_not_evaluated", "gold_answers": false, "reviewer_reference_only": true, "tasks": [ { "id": "WA1", "criteria": [ { "name": "Preserve the existing learning account", "review_question": "The draft still learns the next decision from an expert prefix; it does not invent an unrelated policy or research change.", "evidence_ids": [ "A09" ], "status": "proposed_not_validated" }, { "name": "Make the scientific identity explicit", "review_question": "The proposed formulation is described through the supplied expert-learning mechanism, not an unsupported label or novelty claim.", "evidence_ids": [ "A01", "A09", "A02" ], "status": "proposed_not_validated" }, { "name": "Respect missing comparator evidence", "review_question": "Do not fill the short “graphq” note with unsupported empirical or theoretical advantages.", "evidence_ids": [ "A02" ], "status": "proposed_not_validated" }, { "name": "Separate communication from discovery", "review_question": "Do not imply that the recorded note establishes when imitation learning was invented or implemented.", "evidence_ids": [ "A09", "A02" ], "status": "proposed_not_validated" } ], "later_reference_sources": [ { "evidence_id": "A04", "source_file": "introduction.tex", "content_kind": "manuscript_prose", "text": "Recent work brings learning into SAT but does not fully match the control flow of CDCL. SATformer~\\citep{shi2023satformer} learns instance-level signals and adjusts initial variable activities, then stops acting inside the branching loop. Graph-Q-SAT ~\\citep{kurin2020can} learns a state-centric agent that is queried online, yet it does not directly model the executed branch sequence, and it forgoes clear decision-level supervision. These gaps raise a simple question. Can we learn from an expert that CDCL itself reveals, one that is nearly conflict-free and that reduces propagation by avoiding detours?\n\n\nWe introduce \\emph{[modelname]}, an imitation learner for CDCL branching. From a full run, we collapse backtracks into a short expert keytrace that retains the surviving decisions. Replaying this keytrace on the same instance is nearly conflict-free and removes redundant propagation, which yields clean training targets.\nThen, the learner reads the instance together with a prefix of the keytrace and predicts the next signed variable as the next branch decision.\nAt solve time, the learner acts as a drop-in branching rule under a small query budget, and the solver falls back to the native heuristic when the model is uncertain. \nAll other parts of CDCL remain unchanged, so completeness and robustness are preserved.\n", "text_view": "full_source_excerpt", "presented_text_sha256": "72700532246eb244a7e9adcba26aac21c9ad649f38c3d8d4ae8041df0d603090", "context_status": "A selected manuscript passage, not a complete record of the context available when it was written. The records do not establish who wrote it; scientific claims are unverified.", "excerpt_sha256": "72700532246eb244a7e9adcba26aac21c9ad649f38c3d8d4ae8041df0d603090", "manuscript_revision": 567 }, { "evidence_id": "A10", "source_file": "introduction.tex", "content_kind": "review_macro", "text": "\\reviewmarkA{Classical branching heuristics, however, are hand-crafted and limited in their adaptability. Recent work has sought to improve solver performance by integrating learning-based guidance. For example, SATformer~\\citep{shi2023satformer} learns instance-level signals to adjust variable activities during initialization, but it exerts no influence once the branching loop begins. Graph-Q-SAT~\\citep{kurin2020can} introduces an online agent within CDCL, yet it relies on reinforcement learning (RL), which requires extensive exploration and can be unstable due to sparse rewards and delayed feedback.}", "text_view": "full_source_excerpt", "presented_text_sha256": "83f511c00a12b6fc332f445eae23eccd5e4cd6c8b145dabe52487dce82ccc897", "context_status": "A selected manuscript passage, not a complete record of the context available when it was written. The records do not establish who wrote it; scientific claims are unverified.", "excerpt_sha256": "83f511c00a12b6fc332f445eae23eccd5e4cd6c8b145dabe52487dce82ccc897", "manuscript_revision": 1072 }, { "evidence_id": "A06", "source_file": "introduction.tex", "content_kind": "review_macro", "text": "\\reviewmarkA{In contrast, we adopt imitation learning, which learns directly from expert traces. By training on KeyTraces, i.e., collapsed sequences of surviving decisions from full solver runs, our approach provides clear, dense supervision at every branching step. This allows the learner to reproduce high-quality decisions without costly exploration, yielding faster convergence, more stable training, and a natural alignment with the prefix-conditioned nature of branching.}", "text_view": "full_source_excerpt", "presented_text_sha256": "fbf991cdc3cd81aaac2cab586ce86a8e404891ab9e18cc5cf0046e30c7bcda88", "context_status": "A selected manuscript passage, not a complete record of the context available when it was written. The records do not establish who wrote it; scientific claims are unverified.", "excerpt_sha256": "fbf991cdc3cd81aaac2cab586ce86a8e404891ab9e18cc5cf0046e30c7bcda88", "manuscript_revision": 1072 }, { "evidence_id": "A07", "source_file": "introduction.tex", "content_kind": "manuscript_prose", "text": "We summarize the contributions of this work as below.\n\\begin{itemize}\n \\item We propose ImitSAT, the first branching policy for CDCL solvers based on imitation learning. \nUnlike prior methods that rely on reinforcement learning, ImitaSAT leverages dense, decision-level supervision from expert traces.\n\\item We cast branching as a sequential modeling problem by collapsing solver runs into compact sequences of surviving decisions. These sequences serve as clean, conflict-free training targets and align naturally with prefix-conditioned autoregressive modeling.\n\n\\end{itemize}", "text_view": "full_source_excerpt", "presented_text_sha256": "988a3ad275c539741a6d3711213f91016710a32c859d299ea1010089d4c9d3d3", "context_status": "A selected manuscript passage, not a complete record of the context available when it was written. The records do not establish who wrote it; scientific claims are unverified.", "excerpt_sha256": "988a3ad275c539741a6d3711213f91016710a32c859d299ea1010089d4c9d3d3", "manuscript_revision": 1554 } ], "reference_caution": "Later introduction and review passages may contain unsupported comparative and priority assertions; they are historical alternatives, not approved target answers.", "history_advantage_status": "The supplied history establishes a pre-existing account and a real comparison request. Improvement over a compact current-context prompt is unmeasured.", "comparison_plan": "Compare a source-bounded compact context with the selected historical passages under matched information and length budgets. Do not omit needed technical facts merely to make the history condition win.", "source_assertions_independently_verified": false, "review_status": "proposed_criteria_not_validated", "generated_candidate_output": null, "model_run_performed": false }, { "id": "WA2", "criteria": [ { "name": "Connect the two literature strands", "review_question": "The paragraph identifies their different roles rather than simply appending more citations.", "evidence_ids": [ "B02", "A08", "B04" ], "status": "proposed_not_validated" }, { "name": "Repair logic without changing the mechanism", "review_question": "Motivation connects grammatically to the proposal; the existing technical relationship is preserved.", "evidence_ids": [ "B04" ], "status": "proposed_not_validated" }, { "name": "Control the strength of benefit claims", "review_question": "No added guarantee of speed, low cost, novelty or comparative superiority appears.", "evidence_ids": [ "B04" ], "status": "proposed_not_validated" }, { "name": "Explain the revision", "review_question": "The change note separates language repair from a new scientific contribution and flags evidential uncertainty.", "evidence_ids": [ "B04" ], "status": "proposed_not_validated" } ], "later_reference_sources": [ { "evidence_id": "B05", "source_file": "related works.tex", "content_kind": "manuscript_prose", "text": "Motivated by the limitations of neural guidance for SAT detailed above, and drawing inspiration from imitation learning, we propose \\modelname{}. This branching policy for CDCL clones a near conflict-free KeyTrace distilled from solver runs. To achieve this, we cast branching as prefix-conditioned sequence prediction and train an autoregressive next-decision model on compact sequences of surviving decisions. This provides dense, decision-level supervision at low per-query cost. These design choices help reduce propagation and improve wall-clock time under small query budgets.", "text_view": "full_source_excerpt", "presented_text_sha256": "03c0a3de2d7905c4da6f2cdb3578908d1bc4a8ba460494c1a441cb9f9f9fb637", "context_status": "A selected manuscript passage, not a complete record of the context available when it was written. The records do not establish who wrote it; scientific claims are unverified.", "excerpt_sha256": "03c0a3de2d7905c4da6f2cdb3578908d1bc4a8ba460494c1a441cb9f9f9fb637", "manuscript_revision": 1605 }, { "evidence_id": "B06", "source_file": "related works.tex", "content_kind": "manuscript_prose", "text": "Motivated by the limitations of neural guidance for SAT detailed above, and drawing inspiration from imitation learning, we propose \\modelname{}. The branching policy for CDCL clones a near conflict-free KeyTrace distilled from solver runs. To achieve this, we cast branching as prefix-conditioned sequence prediction and train an autoregressive next-decision model on compact sequences of surviving decisions. This provides dense, decision-level supervision at low per-query cost. These design choices help reduce propagation and improve wall-clock time under small query budgets.", "text_view": "full_source_excerpt", "presented_text_sha256": "fe1cb43071ff2d15c4e227b9509dec27a6072e300b5c99cbd545b47c4326a8cc", "context_status": "A selected manuscript passage, not a complete record of the context available when it was written. The records do not establish who wrote it; scientific claims are unverified.", "excerpt_sha256": "fe1cb43071ff2d15c4e227b9509dec27a6072e300b5c99cbd545b47c4326a8cc", "manuscript_revision": 1716 } ], "reference_caution": "B05 is a recorded later revision, not a gold answer. B06 mostly changes one determiner and should not be treated as a separate high-value outcome.", "history_advantage_status": "The older related-work context may help semantic integration. Local fragment repair is also possible from the current paragraph alone.", "comparison_plan": "Compare current-bridge-only input with the same bridge plus the selected related-work context, controlling for supplied information. Assess integration separately from grammatical repair.", "source_assertions_independently_verified": false, "review_status": "proposed_criteria_not_validated", "generated_candidate_output": null, "model_run_performed": false }, { "id": "WA3", "criteria": [ { "name": "Preserve the definition", "review_question": "The number of model consultations remains clear to a nonspecialist reader.", "evidence_ids": [ "C08" ], "status": "proposed_not_validated" }, { "name": "Separate benefit and cost", "review_question": "Reduced solver work and additional query computation are not conflated.", "evidence_ids": [ "C08" ], "status": "proposed_not_validated" }, { "name": "Remove duplication without losing substance", "review_question": "Repeated active explanations are consolidated; the useful definition and qualifications remain.", "evidence_ids": [ "C08" ], "status": "proposed_not_validated" }, { "name": "Avoid scope regression", "review_question": "The current GQSAT-only runtime subject is not silently broadened to both methods.", "evidence_ids": [ "C08", "C06" ], "status": "proposed_not_validated" }, { "name": "Do not manufacture evidential sufficiency", "review_question": "The proposed method's budget is not called optimal on the basis of an absent figure or comparator-only behavior.", "evidence_ids": [ "C08" ], "status": "proposed_not_validated" } ], "later_reference_sources": [ { "evidence_id": "C09", "source_file": "appendix.tex", "content_kind": "mixed_prose_and_comments", "text": "\\section{Query Budget Analysis}\n\\label{sec:query-budget}\n% This section details our method across different query budgets (1-10) and all model calls.\n% As shown in Figure~\\ref{fig:query}, the performance gain diminishes as the query budget increases. It shows the largest gain from the first 3 queries; later queries, especially after 6 calls, show a smaller gain. \n% We treat the gain from the model, which reduces the propagation number and query times, as a trade-off. So to balance the gain and cost, we set a default query budget of 3. In addition, as shown in Figure~\\ref{fig:wallclock-methods}, both our method and Graph-Q-SAT exhibit worse time performance as the query size increases from 3 to 5.\nIn this section, we study how the query budget, the number of times the solver consults the model during search, affects both effectiveness and computational cost. We vary the budget from 1 to 10 calls, and also include an all-calls configuration, to understand how much benefit each additional model query provides.\n\n% We evaluate our method across different query budgets from 1 to 10 model calls and an all-calls setting.\nAs shown in Figure~\\ref{fig:query}, \nthe performance gain exhibits clear diminishing returns: the majority of improvement is obtained from the first three queries, while additional calls beyond six provide only marginal benefits. Because each query introduces non-trivial latency, we interpret the reduction in propagations achieved by the model as a benefit that must be weighed against the computational overhead of issuing more queries. \nBased on this trade-off, we adopt a default budget of 3 queries. This choice is further supported by the wall-clock results in Figure~\\ref{fig:wallclock-methods}, where GQSAT’s runtime worsens noticeably as the query budget increases from 3 to 5.\n\n% performance exhibits diminishing returns as the query budget increases. Most improvement comes from the first three queries. Additional queries beyond six calls offer only marginal gains.\n% We view the reduction in propagations achieved by the model as a benefit that must be traded off against the computational cost of additional queries.\n% To balance this trade-off, we set a default query budget of 3.\n% Moreover, as shown in Figure~\\ref{fig:wallclock-methods}, GQSAT also exhibit worse wall-clock performance as the query budget increases from 3 to 5.", "text_view": "full_source_excerpt", "presented_text_sha256": "b657f2859c7d6b88481258742093429632df66d1d813a4c23a6e7638b47c91ce", "context_status": "A selected manuscript passage, not a complete record of the context available when it was written. The records do not establish who wrote it; scientific claims are unverified.", "excerpt_sha256": "b657f2859c7d6b88481258742093429632df66d1d813a4c23a6e7638b47c91ce", "manuscript_revision": 2898 } ], "reference_caution": "C09 retains the narrower claim; it does not introduce it. Its stated rationale may still be insufficient to establish the default for the proposed method.", "history_advantage_status": "This is a useful current-draft revision task and a test of harmful reuse from older context. It is not assumed to require history or benefit from more history.", "comparison_plan": "Compare current active C08 alone with C08 plus older active C06. Assess regression as well as clarity; extra history may help, do nothing, or hurt.", "source_assertions_independently_verified": false, "review_status": "proposed_criteria_not_validated", "generated_candidate_output": null, "model_run_performed": false } ], "contributor_privacy": "Private contributor labels use generic markers; scientific wording is preserved.", "use_notice": "For discussing a possible data partnership. Training, redistribution and other uses require a separate agreement." }