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

PLANTAIN: Plan-Answer Interleaved Reasoning

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 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 PLANTAIN, a post-training recipe (synthetic traces, SFT, PPO) that makes Qwen3 models emit a plan first and then interleave thoughts with user-visible answers, plus plan-level Best-of-N and Rewind-and-Repeat inference strategies judged by an LLM autorater.

Problem

Reasoning models produce long hidden thinking before any visible output, which wastes user time when they start from a false premise and offers no chance for early user correction.

Approach

Train models to alternate internal thoughts with short user-facing intermediate responses, starting with an explicit step-by-step plan, then use plan-level rejection sampling or best-of-N selection by an LLM judge to prune flawed plans early before full generation.

Key results

• Plan-Answer on Qwen3-4B cuts average TTFR from 1419 to 516 tokens while raising average P@1 from 56.9 to 58.8 versus Think-Answer.
• R&R (Plan) on Qwen3-8B reaches 64.8 average P@1 with TTFR 571, versus 59.0 and TTFR 1538 for Think-Answer.
• R&R (Answer) reaches similar accuracy but with roughly 6-7x higher TTFR (about 3169-3348 tokens average).

Surprising findings

• Plan-Answer models trained only on coding data generalize to math, text-to-SQL and QuALITY with accuracy gains, not just latency reductions.
• Naively modifying the system instruction without training fails to induce plan-first behavior; models collapse plans into the reasoning trace.

Limitations

• Evaluation uses an LLM autorater as a simulated user rather than real humans.
• Training data is small (125 SFT traces) and focused on structured, verifiable tasks.
• Judge introduces latency and may carry model biases.

Experimental setup

Datasets: BigCodeBench MBPP MATH500 BirdSQL QuALITY

Benchmarks: MATH500 MBPP BirdSQL (Text-to-SQL) QuALITY

Metrics: pass@1 time-to-first-response (thought tokens) tokens-per-problem unit test pass rate

Baselines: No Thinking Think-Answer Rewind-and-Repeat (Answer) Best-of-N (5)

Builds on: Qwen3 PPO GRPO ReAct VERL

Closest papers by meaning

• CodePlan: Unlocking Reasoning Potential in Large Language Models by Scaling Code-form Planning (ICLR 2025 · similarity 0.87)
• Decompose, Analyze and Rethink: Solving Intricate Problems with Human-like Reasoning Cycle (NeurIPS 2024 · similarity 0.86)
• Reasoning Scaffolding: Distilling the Flow of Thought from LLMs (ICLR 2026 · similarity 0.86)
• Plan and Budget: Effective and Efficient Test-Time Scaling on Reasoning Large Language Models (ICLR 2026 · similarity 0.86)
• Reasoning-Trace Collapse: Evaluating the Loss of Explicit Reasoning During Fine-Tuning (NeurIPS 2026 · similarity 0.86)
• SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.86)