Research map / Large language models / Reasoning and chain of thought

ORACLE: Optimizing Reasoning Abilities of Large Language Models via Constraint-Led Synthetic Data Elicitation

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ORACLE, a two-stage synthetic reasoning data framework that uses template-structured LLM steps verified by a symbolic engine (Pyke) inside beam search to build SFT and DPO data.

Problem

Synthetic multi-step reasoning data filtered only by final-answer correctness contains flawed intermediate steps, and existing step verifiers (code execution, symbolic engines) fail on natural language tasks with ambiguous or incomplete context.

Approach

Elicit modular reasoning steps (QUERY, FACTS, RULE, REVISION, REASONING_RESULT) via a fixed template in syllogistic form, then score each candidate step with symbolic engine execution plus LLM precision and feasibility judgments inside beam search; use validated paths for SFT and valid-vs-invalid sibling nodes for DPO.

Key results

• ORACLE on Llama-3.1-8B-Instruct reaches 97.2 on ProntoQA and 67.5 on ProofWriter, exceeding the best baseline (self-rewarding 95.8 and 63.4).
• On Mistral-7B-Instruct-v0.3, ORACLE reaches 87.6 on ProntoQA versus 82.8 for the best baseline (ToT-SFT).
• Ablation shows removing the symbolic engine drops Llama StrategyQA from 91.8 to 89.9.

Surprising findings

• Symbolic engine step execution success is low on commonsense and science tasks (e.g., 24.4% to 30.3% on StrategyQA and ScienceQA for Llama in the first iteration), and in some cases ORACLE's ablations match or beat the full method.

Limitations

• Symbolic engine execution fails often on complex or ambiguous facts and rules, and translation errors by the LLM are common.
• Gains over baselines are sometimes small, and the full method is not best on every dataset-model pair.
• Reported results use only three 7-8B models and the verifier is Pyke.

Experimental setup

Datasets: ProntoQA ProofWriter BoolQ CosmosQA ScienceQA StrategyQA

Benchmarks: ProntoQA ProofWriter BoolQ CosmosQA ScienceQA StrategyQA

Metrics: accuracy step execution success rate

Baselines: CoT zero-shot CoT four-shot RFT ToT-SFT self-rewarding

Builds on: direct preference optimization chain-of-thought rejection sampling fine-tuning tree-of-thoughts self-rewarding language models

Closest papers by meaning

• Enhancing Reasoning Capabilities of LLMs via Principled Synthetic Logic Corpus (NeurIPS 2024 · similarity 0.91)
• Large Language Models Meet Symbolic Provers for Logical Reasoning Evaluation (ICLR 2025 · similarity 0.91)
• Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.90)
• Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling (ICLR 2025 · similarity 0.90)
• Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.90)
• Reasoning-as-Logic-Units: Scaling Test-Time Reasoning in Large Language Models Through Logic Unit Alignment (ICML 2025 · similarity 0.90)