Research map / Large language models / Instruction and fine-tuning
Test-Time Learning for Large Language Models
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 TLM, a test-time learning method that adapts LLMs by minimizing input perplexity on unlabeled test data, using perplexity-based sample selection and LoRA updates, evaluated on the new AdaptEval benchmark.
Problem
LLMs degrade under distribution shifts (domain-specific terminology and linguistic variation) at test time, and existing test-time adaptation methods based on entropy minimization are ill-suited to autoregressive LLMs, while retrieval-based test-time training needs access to training data.
Approach
Minimize the perplexity of unlabeled test inputs (as a proxy for reducing output perplexity), emphasize high-perplexity samples via a perplexity-based selection weight, and update only LoRA parameters to limit catastrophic forgetting.
Key results
- On DomainBench, TLM improves Llama3-8B-Instruct Geography from 0.2450 to 0.3212 and Medicine from 0.1265 to 0.2372 (Rouge-Lsum).
- On ReasoningBench GSM8K, TLM raises Llama3-8B-Instruct from 0.7610 to 0.8074 EM, while EATA drops to 0.0250.
- Sample efficient learning in the online setting reduces backward passes by 69.7% (5000 to 1514) on Llama3-8B-Instruct.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TLM test-time learning instead of the original model | Llama3-8B-Instruct · DomainBench: Geography, Agriculture, Medicine, and Finance | Rouge-Lsum: 0.2450 -> 0.3212; 0.0834 -> 0.1319; 0.1265 -> 0.2372; 0.2329 -> 0.3242 | Helped |
| Use TLM test-time learning instead of the original model | Llama3-8B-Instruct · InstructionBench: Alpaca-GPT4, Dolly, and InstructionWild | Rouge-Lsum: 0.3752 -> 0.4274; 0.3671 -> 0.3785; 0.2608 -> 0.2932 | Helped |
| Use TLM test-time learning instead of the original model | Llama3-8B-Instruct · ReasoningBench: GSM8K, MetaMath, and Logiqa | Exact Match: 0.7610 -> 0.8074; 0.6912 -> 0.7006; 0.4550 -> 0.4868 | Helped |
| Add the Sample Efficient Learning Strategy (ablation) | TLM without SEL · DomainBench: Geography, Agriculture, Medicine, and Finance | Rouge-Lsum: 0.3190 -> 0.3212; 0.1255 -> 0.1319; 0.2326 -> 0.2372; 0.3222 -> 0.3242 | Helped |
| Add the Sample Efficient Learning Strategy (ablation) | TLM without SEL · DomainBench | #Backwards: 5000 -> 4772 | Helped |
| Set perplexity threshold P0 to e^3 rather than other tested thresholds (ablation) | TLM · Geography, Medicine, and Finance | performance | Helped |
| Set perplexity threshold too high or too low (ablation) | TLM · DomainBench | performance | Hurt |
| Use LoRA updates instead of full-parameter updates (ablation) | TLM · GSM8K after adaptation on DomainBench | general-knowledge retention | Helped |
| Train on very low-perplexity samples rather than prioritizing high-perplexity samples | TLM · Test-time learning | performance | Hurt |
Surprising findings
- Entropy-minimization TTA methods (Tent, EATA, COME) frequently degrade LLM performance, sometimes drastically (e.g., EATA on GSM8K for Llama3-8B falls to 0.025).
Limitations
- Gains on the reasoning benchmark are mixed (e.g., Qwen2.5-7B and Llama2-13B show little or negative change on some datasets).
- Theoretical justification relies on an empirical non-negativity assumption on gradient inner products.
Experimental setup
Datasets: GeoSignal Agriculture-QA GenMedGPT-5k Wealth-Alpaca_lora Alpaca-GPT4 Dolly-15k InstructionWild GSM8K MetaMathQA LogiQA
Benchmarks: AdaptEval (DomainBench, InstructionBench, ReasoningBench)
Metrics: Rouge-Lsum Exact Match
Baselines: Tent EATA COME
Builds on: Tent EATA COME LoRA Test-Time Training
Closest papers by meaning
- In-Place Test-Time Training (ICLR 2026 · similarity 0.90)
- Model Whisper: Steering Vectors Unlock Large Language Models’ Potential in Test-Time (AAAI 2026 · similarity 0.90)
- Compute Aligned Training: Optimizing for Test Time Inference (NeurIPS 2026 · similarity 0.89)
- LensLLM: Unveiling Fine-Tuning Dynamics for LLM Selection (ICML 2025 · similarity 0.88)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.88)
- Inference Scaling Laws: An Empirical Analysis of Compute-Optimal Inference for LLM Problem-Solving (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.