{
  "schema": "project-experience.writing-task-input.v1",
  "status": "constructed_example_not_evaluated",
  "context_rule": "Only supplied excerpts with Overleaf revision <= cutoff are included. Prompts and selected cutoffs were designed retrospectively; this is not original-context replay or a sealed evaluation.",
  "task": {
    "id": "WA2",
    "title": "Connect literature strands without rewriting the science",
    "action": "Draft an argument bridge",
    "cutoff_revision": 1597,
    "prompt": "Revise the supplied draft bridge so a reader can follow the connection from prior solver guidance and imitation learning to the proposal. Use only the supplied manuscript passages; their descriptions of prior work and performance are not independently verified. Repair the opening fragment, separate motivation from mechanism, and keep the next-decision supervision account intact. Do not add citations, strengthen benefit claims or invent a new experimental result. Put any unsupported-benefit concern in a separate short review note.",
    "output_request": "One revised paragraph and a short explanation of what was preserved, changed and still needs evidence. These instructions are newly authored.",
    "skill_ids": [
      "WS02",
      "WS03",
      "WS07",
      "WS05"
    ],
    "input_note": "The current bridge is paired with selected older passages from the same related-work file, not a complete file snapshot at the cutoff.",
    "current_source_ids": [
      "B04"
    ],
    "earlier_context_ids": [
      "B02",
      "A08"
    ],
    "sources": [
      {
        "evidence_id": "B04",
        "source_file": "related works.tex",
        "content_kind": "manuscript_prose",
        "text": "Based on the above limitation with neural guidance for SAT, and inspired by imitation learning. We propose \\modelname{}, a branching policy for CDCL that clones a near conflict‑free KeyTrace distilled from solver runs. We cast branching as prefix-conditioned sequence prediction and train an autoregressive next-decision model on compact sequences of surviving decisions, providing dense, decision-level supervision with low per-query cost. These designs help reduce propagation and improve wall-clock time under small query budgets.",
        "text_view": "full_source_excerpt",
        "presented_text_sha256": "d3bd52fcc95805ef9a0bd17792caa3651a840a5d4964738529325687d26e5915",
        "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": "d3bd52fcc95805ef9a0bd17792caa3651a840a5d4964738529325687d26e5915",
        "manuscript_revision": 1597
      },
      {
        "evidence_id": "B02",
        "source_file": "related works.tex",
        "content_kind": "manuscript_prose",
        "text": "\\section{Related Works}\n\n\\paragraph{Neural guidance for SAT and CDCL.}\nEarly learning approaches focused on instance-level prediction, using Graph Neural Network (GNNs)~\\citep{scarselli2008graph} to classify SAT or UNSAT, as seen in NeuroSAT~\\citep{selsam2018learning,selsam2019guiding} and ~\\citep{cameron2020predicting}. Recent work has explored whether Transformers can learn solver behavior directly~\\citep{pan2025can}.\nIn parallel with these model-based approaches, complementary efforts target data and benchmarking, including G2SAT~\\citep{you2019g2sat} and G4SATBench~\\citep{lig4satbench}.\nBuilding on these foundations, a second line integrates learning inside solvers to shape specific components: for example, NeuroSelect~\\citep{liu2024neuroselect} learns clause deletion policies, NeuroBack~\\citep{wang2024neuroback} improves phase initialization with GNNs, and RDC‑SAT~\\citep{zhai2025learning} adopts a divide-and-conquer strategy via reinforcement learning. This approach leads to targeted enhancements within solver mechanisms.\nMore concretely, within the CDCL branching loop, several methods exemplify this integration: NeuroSAT~\\citep{selsam2019guiding} has been used to guide variable selection; Graph-Q-SAT~\\citep {kurin2020can} trains an RL agent queried online during search based on instance information; and SATformer~\\citep{shi2023satformer} trains a GNN Transformer model to initialize the CDCL that indirectly influences branching thereafter.\n\n\n\\paragraph{Imitation learning.}",
        "text_view": "full_source_excerpt",
        "presented_text_sha256": "214a40d95caf6e9de312dce02260293e33fb53b3aaec31906f4534399b08ffeb",
        "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": "214a40d95caf6e9de312dce02260293e33fb53b3aaec31906f4534399b08ffeb",
        "manuscript_revision": 1576
      },
      {
        "evidence_id": "A08",
        "source_file": "related works.tex",
        "content_kind": "manuscript_prose",
        "text": "\\paragraph{Imitation learning for control.}\nImitation learning (IL) learns policies directly from expert demonstrations, that is, sequences of states with associated actions~\\citep{osa2018algorithmic,zare2024survey}. A simple example is behavior cloning (BC), which utilizes supervised learning to map observed situations to expert choices~\\citep{pomerleau1991efficient}.\nBuilding on the principles of imitation learning, the Decision Transformer~\\citep{chen2021decision} is similar to behavior cloning, framing reinforcement learning as sequence modeling, where an autoregressive Transformer is trained to predict the next action given a sequence rollout of returns, states, and actions. This view connects control to next-token prediction and attains competitive performance without explicit value function learning.\nThe application of imitation learning extends beyond traditional domains. For instance, beyond robotics and games, IL has guided decision-making in exact optimization solvers. In mixed-integer linear programming, policies learned to imitate strong branching can be used within branch-and-bound and achieve strong results~\\citep{gasse2019exact}. Related work also learns branching policies that integrate into branch-and-bound~\\citep{zarpellon2021parameterizing}.",
        "text_view": "full_source_excerpt",
        "presented_text_sha256": "241387f79147cf211eb2162a15051cb4946a63ed85f769293b9f0b4c1ed37670",
        "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": "241387f79147cf211eb2162a15051cb4946a63ed85f769293b9f0b4c1ed37670",
        "manuscript_revision": 1585
      }
    ],
    "prompt_origin": "newly_authored_derived_instruction_not_historical_prompt",
    "historical_note_available": false,
    "reference_completion_included": false,
    "independent_project_group": "ImitSAT-single-project"
  },
  "version": "1.0",
  "use_notice": "For discussing a possible data partnership. Training, redistribution and other uses require a separate agreement."
}