The Ledger
Every row is one claim: a technique, on a task type, for a model family.
| Technique | Task type | Model families | Status | Effect | Source |
|---|---|---|---|---|---|
| few-shot | domain-multimodal | gpt-4o | recommended | 52.17% zero-shot to 73.52% at 16-shot | arXiv:2509.09307 |
| few-shot | output-formatting | all | recommended | Still the cheapest way to pin an output schema | arXiv:2506.14641 |
| few-shot-cot | reasoning | qwen2.5 | redundant | No improvement over zero-shot CoT | arXiv:2506.14641 |
| self-consistency | reasoning | lamda-pt, gpt-3 | recommended | majority-voting over sampled chain-of-thought reasoning paths improves reasoning accuracy over single-path greedy decoding | arXiv:2203.11171 |
| self-critique | code-generation | codegen, starcoder | recommended | Fine-tuning code LMs to self-refine faulty generations with execution feedback lifts pass rates by up to 63.5% relative | arXiv:2403.18746 |
| step-back | reasoning | palm-2l | recommended | Substantial accuracy gains for PaLM-2L on MMLU STEM, TimeQA, MuSiQue, and StrategyQA reasoning benchmarks | arXiv:2310.06117 |
| structured-output | output-formatting | dream-7b, llada-8b | recommended | FA-constrained decoding substantially improves accuracy on schema-constrained generation | arXiv:2607.07026 |
| structured-output | output-formatting | llama-3.1 | recommended | GCD/P-GCD proposals lift constraint satisfaction from 95.26–99.20% (LCD) to 100.0% on structured JSON/Python function-calling outputs | arXiv:2606.01926 |
| take-a-deep-breath | reasoning | all | unverified | Widely repeated; no supporting study located |