Method Drift

Which methods have been superseded

A living systematic review per builder-problem. For each problem below, we track which methods have become superseded baselines, what sits on the live frontier, and the grounded receipts — verbatim critique quotes and benchmark numbers, each traced to the paper it came from — behind every “X superseded by Y” claim.

10 builder-problems · 3,164 papers · 6,756 critique receipts · 26,480 benchmark results

Parameter-efficient fine-tuning (LoRA family)

Adapting big models without touching every weight — LoRA and the variants chasing it.

721 papers · 1,563 receiptsTop baseline: LoRA

Retrieval-augmented generation

Grounding generation in retrieved context — the architectures replacing naive RAG.

664 papers · 1,433 receiptsTop baseline: RAG

LLM reasoning / chain-of-thought

Teaching models to think before answering — chain-of-thought and what came after.

391 papers · 892 receiptsTop baseline: Chain-of-Thought

Mixture-of-experts routing

Scaling capacity without the compute — sparse experts and how they are routed.

655 papers · 1,456 receiptsTop baseline: MC-SMoE

KV-cache compression

Shrinking the attention KV cache for long context — eviction, quantization, and reuse.

264 papers · 613 receiptsTop baseline: SnapKV

Speculative decoding

Drafting tokens ahead to speed up inference — the EAGLE/Medusa lineage.

182 papers · 333 receiptsTop baseline: EAGLE-3

LLM quantization

Running large models in low precision — post-training quantization methods.

97 papers · 151 receiptsTop baseline: GPTQ

Agent / long-term memory

What an agent remembers across a session — long-term memory systems.

69 papers · 99 receiptsTop baseline: Mem0

Long-context / context-window extension

Reading past the training length — positional tricks that extend context.

69 papers · 137 receiptsTop baseline: RoPE

Tool use / function calling

Teaching models to call APIs and tools — function-calling and agent toolkits.

52 papers · 79 receiptsTop baseline: ReAct