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

Algorithm Generation via Creative Ideation

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning

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 MetaMuse, a creative ideation framework for LLM-driven system algorithm generation that grounds diversity in performance feedback, steers ideation with random dictionary keywords, and uses waypoint reasoning to produce executable cache-replacement and bin-packing heuristics.

Problem

LLMs generating system algorithms are biased toward well-known generic heuristics (availability bias), failing to make the discontinuous leaps needed for high-performing designs in cache replacement and online bin packing.

Approach

Self-reflection principles: measure solution diversity in feedback (performance) space rather than idea space, steer ideation with external keyword stimuli instead of temperature randomness, and build executable solutions through structured waypoints (property extraction, problem mapping, solution formulation, code generation) instead of free-form CoT; RSDict-SF uses Gaussian process regression ov

Key results

• On cache replacement, MetaMuse reduces miss ratio by up to 35.76% vs human heuristics and up to 9.89% vs LLM-based baselines (GPT-4o, 90th/75th percentile traces).
• On online bin packing with DeepSeek-V3, MetaMuse reduces bin usage by up to 30.93% vs human heuristics and up to 21.06% vs LLM-based baselines.
• MetaMuse generates up to 1.78x more distinct cache replacement solutions and 1.80x more distinct bin packing solutions than LLM-based baselines, at about 2.16 cents per solution with GPT-4o.

Surprising findings

• Raising LLM temperature only marginally mitigates availability bias, since it is a monotonic transformation that preserves token rankings.
• Making Repeated Sampling aware of previously generated solutions does not reduce the bias toward classical heuristics like LRU, LFU, and FIFO.

Limitations

• Stimuli are restricted to a general English keyword dictionary; other stimulus sources are unexplored.
• RSDict-SF requires benchmarking every solution on n workload traces, which can be costly for end-to-end systems.
• Solutions are evaluated in simulation on traces; production deployment is still ongoing.

Experimental setup

Datasets: RetrievalAttention traces Tencent block storage traces Alibaba cloud storage traces BPPLIB Falkenauer-U and Scholl-1 Weibull synthetic traces Gaussian synthetic traces libCacheSim synthetic Zipfian traces

Benchmarks: RetrievalAttention cache traces Tencent block storage cache traces Alibaba cloud storage cache traces BPPLIB bin packing instances Weibull bin packing traces Gaussian bin packing traces

Metrics: cache miss ratio bin usage number of distinct solutions cluster density cost per solution (cents)

Baselines: MCTS-AHD ReEvo OpenEvolve PlanSearch Repeated Sampling LRU LFU FIFO Sieve S3FIFO TinyLFU ARC

Builds on: PlanSearch Repeated Sampling libCacheSim

Closest papers by meaning

• Chain of Thoughtlessness? An Analysis of CoT in Planning (NeurIPS 2024 · similarity 0.84)
• PathWise: Planning through World Model for Automated Heuristic Design via Self-Evolving LLMs (ICML 2026 · similarity 0.84)
• JADE: Bridging the Strategic-Operational Gap in Dynamic Agentic RAG (ICML 2026 · similarity 0.83)
• Algorithm of Thoughts: Enhancing Exploration of Ideas in Large Language Models (ICML 2024 · similarity 0.83)
• Eco Search: A No-delay Best-First Search Algorithm for Program Synthesis (AAAI 2025 · similarity 0.83)
• Thought of Search: Planning with Language Models Through The Lens of Efficiency (NeurIPS 2024 · similarity 0.83)