{ "id": 615162, "title": "VishnuDharma ARC Solver: A Hybrid AI Framework", "content": "

🧩 Overview of Our Approach

\n

To solve ARC (Abstraction and Reasoning Corpus) tasks, our team developed a Hybrid Reasoning Framework that combines:

\n\n

This blend enables our method to perform symbolic reasoning, adapt to novel problems, and stay fully compatible with Kaggle's offline constraints.

\n

🎯 Key Design Goals

\n\n

πŸ› οΈ Solution Architecture

\n

We used a 3-stage cascaded process:

\n

Stage 1 – Heuristic Solver

\n\n

Stage 2 – LLM Supervisor (Optional)

\n\n

Stage 3 – Beam Search Program Synthesis

\n\n

βœ” If all three stages fail, we use a controlled fallback strategy, like:

\n\n

πŸ“Š Workflow

\n
    \n
  1. Load ARC task.
    2. \n
  2. Extract colors, objects, shapes, size differences.
    3. \n
  3. Try basic heuristic transformations.
    4. \n
  4. If unsure β†’ ask LLM: β€œIs this rotation, reflection or color change?”
    5. \n
  5. If still unsolved β†’ run beam search over DSL functions.
    6. \n
  6. Validate the function on training examples.
    7. \n
  7. Apply to test inputs.
    8. \n
  8. Save two attempts per task in submission.json.
    9. \n
\n

🧠 Why This Works

\n\n

πŸ“ Output Format (submission.json Example)

\n

{\n \"00576224\": [\n {\n \"attempt_1\": [[0,1],[1,0]],\n \"attempt_2\": [[0,0],[0,0]]\n }\n ]

\n

}

\n

πŸ‘₯ Team Members – VishnuDharma

\n", "comments": [], "comment_count": 0, "messages": [ { "id": 3313490, "postDate": "2025-11-09T10:49:17.277Z", "content": "

🧩 Overview of Our Approach

\n

To solve ARC (Abstraction and Reasoning Corpus) tasks, our team developed a Hybrid Reasoning Framework that combines:

\n\n

This blend enables our method to perform symbolic reasoning, adapt to novel problems, and stay fully compatible with Kaggle's offline constraints.

\n

🎯 Key Design Goals

\n\n

πŸ› οΈ Solution Architecture

\n

We used a 3-stage cascaded process:

\n

Stage 1 – Heuristic Solver

\n\n

Stage 2 – LLM Supervisor (Optional)

\n\n

Stage 3 – Beam Search Program Synthesis

\n\n

βœ” If all three stages fail, we use a controlled fallback strategy, like:

\n\n

πŸ“Š Workflow

\n
    \n
  1. Load ARC task.
    2. \n
  2. Extract colors, objects, shapes, size differences.
    3. \n
  3. Try basic heuristic transformations.
    4. \n
  4. If unsure β†’ ask LLM: β€œIs this rotation, reflection or color change?”
    5. \n
  5. If still unsolved β†’ run beam search over DSL functions.
    6. \n
  6. Validate the function on training examples.
    7. \n
  7. Apply to test inputs.
    8. \n
  8. Save two attempts per task in submission.json.
    9. \n
\n

🧠 Why This Works

\n\n

πŸ“ Output Format (submission.json Example)

\n

{\n \"00576224\": [\n {\n \"attempt_1\": [[0,1],[1,0]],\n \"attempt_2\": [[0,0],[0,0]]\n }\n ]

\n

}

\n

πŸ‘₯ Team Members – VishnuDharma

\n", "rawMarkdown": "🧩 Overview of Our Approach\n\nTo solve ARC (Abstraction and Reasoning Corpus) tasks, our team developed a Hybrid Reasoning Framework that combines:\n - Heuristic rule discovery,\n - Lightweight LLM-based supervision, and\n - Program synthesis via beam search,\n along with a fallback system to ensure valid output for every task.\n\nThis blend enables our method to perform symbolic reasoning, adapt to novel problems, and stay fully compatible with Kaggle's offline constraints.\n\n🎯 Key Design Goals\n - No reliance on internet or cloud models\n - Explainable solutions rather than black-box predictions\n - Generalization to unseen ARC tasks\n - Every task should produce a valid submission (no task left blank)\n\nπŸ› οΈ Solution Architecture\n\n\nWe used a 3-stage cascaded process:\n\nStage 1 – Heuristic Solver\n - Description: Detects direct transformations: color maps, flips, rotations, translations, cropping, repetition.\n - When Used: First attempt\n\nStage 2 – LLM Supervisor (Optional)\n - Description: Local LLM like LLaMA/FlanT5 predicts \"what kind of transformation\" (e.g., reflect, recolor, move)\n - When Used: If heuristics are unsure\n\nStage 3 – Beam Search Program Synthesis\n - Description: Composes small functions (rotate β†’ crop β†’ recolor) using beam search to find a program f such that f(X_train) = Y_train\n - When Used: If LLM or heuristics fail\n\nβœ” If all three stages fail, we use a controlled fallback strategy, like:\n - Copy input β†’ output\n - Most frequent color fill\n - Identity mapping to avoid invalid submission.json\n\nπŸ“Š Workflow\n1. Load ARC task.\n2. Extract colors, objects, shapes, size differences.\n3. Try basic heuristic transformations.\n4. If unsure β†’ ask LLM: β€œIs this rotation, reflection or color change?”\n5. If still unsolved β†’ run beam search over DSL functions.\n6. Validate the function on training examples.\n7. Apply to test inputs.\n8. Save two attempts per task in submission.json.\n\n🧠 Why This Works\n - Heuristics solve easy tasks quickly; beam search solves complex ones.\n - Not dependent on past data; works on unseen problems.\n - Inspired by human reasoning, compositionality, and ARC principles from FranΓ§ois Chollet.\n - Every stage is explainable and reproducible offline.\n - Our system blends symbolic reasoning + LLM judgment + fallback safety.\n - This hybrid approach brings ARC closer to human-like reasoning and pushes AGI-inspired problem-solving.\n\nπŸ“ Output Format (submission.json Example)\n==========\n{\n \"00576224\": [\n {\n \"attempt_1\": [[0,1],[1,0]],\n \"attempt_2\": [[0,0],[0,0]]\n }\n ]\n}\n==========\n\nπŸ‘₯ Team Members – VishnuDharma\n - Pranjul Tiwari – AI/LLM Systems & Program Synthesis\n - Ankit Tiwari – Strategy, Optimization & Solution Architecture\n\n" } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T21:00:15Z", "competition_id": 91496, "competition_slug": "arc-prize-2025", "team_id": 14457891, "solution_url": "/competitions/arc-prize-2025/writeups/vishnudharma-arc-solver-a-hybrid-ai-framework" }