Overview
Production Readiness
0.4
Novelty Score
0.7
Cost Impact Score
0.6
Citation Count
0
Why It Matters For Business
Automating algorithm discovery with LLMs can find improved model-compression strategies and reduce expert hours needed to hand-craft search heuristics.
Summary TLDR
tnGPS is a prompt-driven pipeline that uses large language models to generate new tensor-network-structure-search (TN-SS) algorithms, runs them in a sandboxed evaluator, and iterates. On image compression and Gaussian-process model-compression benchmarks the discovered algorithms beat several prior heuristics; an ablation shows the LLM-driven steps (recombine, increment, inject diversity) matter. Results vary with LLM choice and require careful evaluation.
Problem Statement
Choosing tensor-network topology and ranks is a high-dimensional discrete search. Existing TN-SS heuristics are hand-designed, often trade off exploration vs exploitation poorly, and can get stuck or need heavy expert work. The paper asks if LLMs can automate discovery of better TN-SS algorithms.
Main Contribution
A practical framework, tnGPS, that uses LLMs (prompt pipeline) to generate, refine, and diversify TN-SS sampling functions and then evaluates them automatically.
Empirical evidence that tnGPS can produce new TN-SS algorithms that outperform several state-of-the-art heuristics on in-domain (image compression) and out-of-domain (Gaussian-process compression) benchmarks.
Key Findings
tnGPS produced best-found algorithm with a much lower objective value than baselines on the reported benchmark.
A discovered algorithm (Ho-2) showed higher compression on held-out images than prior methods.
Ablation of LLM-driven phases hurts final performance.
Stronger LLMs give better discovered algorithms; differences are present but modest in tested setup.
Results
Objective (lower better)
Log compression ratio (higher better)
Objective across LLMs (lower better)
Who Should Care
What To Try In 7 Days
Clone the tnGPS repo and run the sandbox on a small image (BSD500 subset) to reproduce Ho-2.
Use GPT-4 with the provided prompts to generate a few GenerateSample variants and run them locally.
Run the ablation (disable KR/II/DI) to see how each prompt stage affects results on your data.
Agent Features
Memory
- history_populations (past sampled algorithms and scores) used as context
Planning
- iterative generate-evaluate-refine loop
- bi-level roulette selection for idea dropout
Tool Use
- LLM code generation
- sandboxed code execution for evaluation
Frameworks
- tnGPS
Is Agentic
true
Architectures
- prompt-pipeline
Collaboration
- LLM uses in-context examples (existing algorithms) as meta-prompts
Reproducibility
Data Urls
- BSD500 (public), CCPP (public), MG (public)
Code Available
Data Available
Open Source Status
- partial
Risks & Boundaries
Limitations
- Final algorithm quality depends on the LLM used; results vary across models.
- Requires nontrivial compute to evaluate many generated algorithms.
- LLMs can hallucinate runnable code; sandbox checks help but may not catch subtle bugs.
- Discovery is driven by the initial pool of algorithms and prompt design, which can bias outcomes.
When Not To Use
- When you need provable algorithmic guarantees or formal correctness.
- If you lack compute to evaluate many candidate algorithms.
- When data or code cannot be exposed to external LLM APIs due to privacy.
Failure Modes
- LLM produces plausible-looking but incorrect or inefficient code that escapes sandbox checks.
- Overfitting discovered heuristics to training images or a narrow dataset.
- Path dependence: generated ideas remain close to the initial pool and miss fundamentally different tactics.
Core Entities
Models
- gpt-4-1106-preview
- gpt-3.5-turbo-16k-0613
- claude-instant-1
- claude-2
- TNGA
- TNLS
- GREEDY
- TnALE
- Ho-1
- Ho-2
- Ho-3
- tnGPS
Metrics
- Objective (Eq.1 combining complexity and relative squared error)
- Log compression ratio
- Relative squared error (RSE)
Datasets
- BSD500 (subset)
- CCPP
- MG
Benchmarks
- Image compression (tensor decomposition)
- Model compression for tensorial Gaussian Process