Research map / Large language models / Instruction and fine-tuning

Improving Complex Reasoning with Dynamic Prompt Corruption: A Soft Prompt Optimization Approach

ICLR 2025 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data

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 Dynamic Prompt Corruption (DPC), an instance-level method that detects soft prompts harming CoT reasoning via saliency-based information-flow analysis and then masks the most influential soft prompt token, improving prompt tuning on GSM8K, MATH, and AQuA.

Problem

Prompt tuning of frozen LLMs gives little gain and can even degrade performance on complex multi-step reasoning tasks, because soft prompts help some instances and harm others.

Approach

Uses saliency scores (attention times gradient) to detect when soft prompts accumulate information and disrupt late-stage reasoning flow, then applies instance-level corruption (masking the most accumulated soft prompt vector plus sparse masking of small embedding values) only for flagged instances.

Key results

• DPC raises GSM8K accuracy for LLaMA3-8B from 65.5% (prompt tuning) to 67.6%, and AQuA from 38.5% to 42.5%.
• On LLaMA2-13B, DPC improves AQuA from 22.4% to 31.1% and GSM8K from 38.1% to 41.9%.
• Random corruption drops LLaMA3-8B GSM8K accuracy to 51.3%, below prompt tuning's 65.5%.

Measured effects reported in the paper

Surprising findings

• Prompt tuning gives only a 0.6% gain over the pretrained LLaMA3-8B on GSM8K, and can degrade performance on some tasks.
• Corruption applied to all instances without the dynamic trigger yields only marginal gains, and random corruption hurts performance substantially.

Limitations

• Identifying erroneous flows requires a complete reasoning pass, increasing inference cost.
• Analysis may not be a universal interpretation; the observed correlations are not confirmed in every case.
• Thresholds (alpha, beta, Gamma) were chosen heuristically after comparative analysis of instances.

Experimental setup

Datasets: GSM8K MATH AQuA

Benchmarks: GSM8K MATH AQuA ProntoQA ARC-challenge WorldTree

Metrics: accuracy

Baselines: Prompt tuning ACT Pretrained model Random corruption

Builds on: prompt tuning chain-of-thought attention calibration technique (ACT) StreamingLLM

Closest papers by meaning

• Revisiting Prefix-tuning: Statistical Benefits of Reparameterization among Prompts (ICLR 2025 · similarity 0.87)
• Scaling Prompt Synthesis for Large Language Model Reasoning (ICML 2026 · similarity 0.87)
• From Noise to Diversity: Random Embedding Injection in LLM Reasoning (NeurIPS 2026 · similarity 0.87)
• How to Auto-optimize Prompts for Domain Tasks? Adaptive Prompting and Reasoning through Evolutionary Domain Knowledge Adaptation (NeurIPS 2025 · similarity 0.87)
• Eliciting Causal Abilities in Large Language Models for Reasoning Tasks (AAAI 2025 · similarity 0.87)
• GReaTer: Gradients Over Reasoning Makes Smaller Language Models Strong Prompt Optimizers (ICLR 2025 · similarity 0.87)