Precedent-Informed Reasoning: Mitigating Overthinking in Large Reasoning Models via Test-Time Precedent Learning

1. Problem Statement

Large Reasoning Models (LRMs), such as OpenAI's o1 and DeepSeek-R1, excel at complex tasks by generating long chains of thought (CoT). However, they frequently suffer from "overthinking":

• Redundant Self-Exploration: Models engage in excessive trial-and-error, re-validating intermediate steps, and exploring irrelevant solution paths.
• Inefficiency: This behavior inflates computational costs (token generation) and inference latency.
• Performance Degradation: In some cases, over-commitment to exploration leads to lower accuracy.
• Limitation of Existing Solutions: Current methods either rely on multi-model collaboration (high resource cost) or test-time control strategies (e.g., early stopping) that merely truncate reasoning without altering the underlying inefficient search paradigm. They fail to leverage past successful solutions to guide current reasoning.

The authors propose mimicking human reasoning, where new problems are solved by recalling and adapting precedents (similar past cases) to constrain the search space, rather than starting from scratch.

2. Methodology: Precedent-Informed Reasoning (PIR)

PIR is a test-time framework that transforms the reasoning paradigm from exhaustive self-exploration to guided learning from precedents. It consists of two core components:

A. Adaptive Precedent Selection (APS)

Goal: Determine "What to imitate?"
Instead of using a fixed number of few-shot examples, APS dynamically selects a compact, high-quality set of precedents for each specific query and model.

1. Semantic Similarity: Candidates are ranked by cosine similarity between the query and precedent embeddings.
2. Model-Specific Informativeness: Candidates are further scored based on the perplexity of the precedent question under the target LRM. Lower perplexity indicates the model is already familiar with the pattern, making it a better candidate for internalization.
3. Adaptive Quantity Determination: Inspired by prequential evaluation, the system tests different numbers of precedents ($m$) to find the count that maximizes the perplexity reduction of the resulting instruction. This ensures the selected precedents provide the maximum "information gain" for the specific model-task pair.

B. Test-Time Experience Internalization (TEI)

Goal: Determine "How to imitate?"
Simply appending precedents to the prompt (passive In-Context Learning) is insufficient. TEI actively forces the model to internalize the solution patterns.

1. Instruction Construction: A "precedent-informed instruction" is constructed by concatenating the target question with the selected precedents.
2. Lightweight Adaptation: The model undergoes test-time learning using LoRA (Low-Rank Adaptation). The base parameters are frozen; only the LoRA adapters are updated for a single step (or very few steps) to minimize the perplexity of the constructed instruction.
3. Behavioral Prior: This update effectively "teaches" the model the specific reasoning pattern of the precedents as a prior. The adapted model then generates the final answer, constrained by this internalized pattern, significantly reducing the need for redundant exploration.

3. Key Contributions

1. Paradigm Shift: Proposes a shift from "self-exploration" to "precedent-informed reasoning," aligning LRM behavior with human problem-solving strategies.
2. Dual-Stage Framework: Introduces APS for model-aware, adaptive precedent selection and TEI for active, label-free test-time internalization of solution patterns.
3. Efficiency-Accuracy Trade-off: Demonstrates that PIR achieves superior accuracy-efficiency trade-offs, consistently shortening reasoning traces while maintaining or improving accuracy across diverse tasks.
4. Empirical Validation: Extensive experiments on mathematical reasoning, scientific Q&A, and code generation benchmarks.

4. Experimental Results

The authors evaluated PIR on DeepSeek-R1-Distilled (7B/14B/32B) and QwQ-32B across five datasets: GSM8K, MATH500, AMC AIME, GPQA Diamond, and LiveCodeBench.

• Accuracy vs. Length:
  • On LiveCodeBench, PIR improved accuracy by 11.61% while reducing reasoning length by 28.30% compared to the second-best method.
  • On AMC AIME and GPQA Diamond, PIR achieved the highest accuracy among all baselines with the shortest reasoning traces.
• Comparison with Baselines:
  • Outperformed training-free methods (e.g., DEER, SEAL) and fine-tuned methods (e.g., s1, L1).
  • Significantly outperformed multi-model collaboration methods (e.g., Speculative Thinking) in terms of latency and resource usage.
• Ablation Studies:
  • Removing APS (using fixed precedents) degraded performance.
  • Removing TEI (using only prompting without LoRA adaptation) resulted in longer reasoning traces and lower accuracy, proving that active internalization is critical.
• Attention Analysis:
  • PIR showed a higher ratio of "high-attention" tokens and segments compared to baselines, indicating that the model focuses on relevant information rather than redundant exploration.
• Cost:
  • The test-time adaptation adds negligible latency (approx. 1-1.5 seconds) but reduces total inference time by up to 3.19x on complex tasks due to the drastic reduction in generated tokens.

5. Significance

• Scalability: PIR is model-agnostic and scales effectively across different model sizes (7B to 32B+).
• Practicality: It offers a deployable solution for reducing inference costs in production LLMs without requiring expensive retraining or multi-GPU setups.
• Theoretical Insight: The paper provides evidence that "overthinking" in LRMs is a solvable inefficiency that can be mitigated by leveraging external knowledge (precedents) and active parameter adaptation, rather than just controlling generation length.
• Future Direction: It highlights the potential of test-time learning as a viable strategy for real-time reasoning optimization, moving beyond static prompting or heavy fine-tuning.

In conclusion, Precedent-Informed Reasoning (PIR) effectively curbs the "overthinking" phenomenon in Large Reasoning Models by dynamically selecting relevant past cases and actively internalizing their solution patterns, resulting in faster, cheaper, and more accurate reasoning.