Skip to content

Latest commit

 

History

History
65 lines (55 loc) · 7.91 KB

File metadata and controls

65 lines (55 loc) · 7.91 KB

Wave 3+ D' — Hybrid Scale Benchmark

  • Manifest: 200 entries (18 real + 182 junk)
  • Queries: 108
  • Runs per (pipeline × query): 3
  • Pipelines: llm-only (deepseek-v3.2 over full manifest) vs hybrid (BM25 top-50 → deepseek-v3.2)
  • Total calls: 648

Aggregated (all runs flattened)

Pipeline Top-1 Top-5 p50 latency p90 latency Avg manifest handed to LLM
llm-only 249/324 (76.9%) 289/324 (89.2%) 2056ms 3595ms 200
hybrid 257/324 (79.3%) 287/324 (88.6%) 1509ms 2725ms 50

Per-run top-1 accuracy (variance inspection)

Pipeline run 0 run 1 run 2 mean stddev
llm-only 76.9% 76.9% 76.9% 76.85% 0.00 pp
hybrid 78.7% 79.6% 79.6% 79.32% 0.44 pp

Per-query run-to-run consistency

Pipeline Queries w/ identical top-1 across all runs Partial disagreement
llm-only 100/108 8/108
hybrid 100/108 8/108

Top-1 disagreements between pipelines (any run)

  • Q9 "What are the stats for AI Day 2026 event attendance and sess" ideal=20260411-garnet-ai-digest-division-of-it-at-usc.md llm→20260411-garnet-ai-digest-division-of-it-at-usc.md hybrid→20260411-garnet-ai-digest-division-of-it-at-usc.md
  • Q17 "I'm not looking for AI benefits, but risks of autonomous AI " ideal=20260411-2510-05179v2.md llm→20260411-evensong-paper-zh.md hybrid→20260411-evensong-paper-zh.md
  • Q30 "What are innovative memory models for AI conversation system" ideal=20260411-2604-08256-hypermem.md llm→20260411-2604-08256-hypermem.md hybrid→20260411-2604-08256-hypermem.md
  • Q36 "What are the effects of memory systems on autonomous technol" ideal=20260411-evensong-paper-zh.md llm→20260411-2604-08256-hypermem.md hybrid→20260411-2310-08560-memgpt.md
  • Q38 "What methods allow AI agents to learn from past mistakes usi" ideal=20260411-2303-11366-reflexion.md llm→20260411-2303-11366-reflexion.md hybrid→20260411-evensong-paper-zh.md
  • Q39 "What are the results of language agents on HumanEval Python " ideal=20260411-2303-11366-reflexion.md llm→20260411-2303-11366-reflexion.md hybrid→20260411-2303-11366-reflexion.md
  • Q42 "Can AI systems get better at coding tasks through memory tec" ideal=20260411-2303-11366-reflexion.md llm→20260411-2604-08256-hypermem.md hybrid→20260411-evensong-paper-zh.md
  • Q44 "What approach helps language models manage memory across mul" ideal=20260411-2310-08560-memgpt.md llm→20260411-2310-08560-memgpt.md hybrid→20260411-msa-memory-sparse-attention.md
  • Q47 "I'm not looking for basic AI chatbots, but advanced memory m" ideal=20260411-2310-08560-memgpt.md llm→20260411-2604-08256-hypermem.md hybrid→20260411-competitor-landscape.md
  • Q48 "What are innovative ways to handle memory in AI systems?" ideal=20260411-2310-08560-memgpt.md llm→20260411-2604-08256-hypermem.md hybrid→20260411-2604-08256-hypermem.md
  • Q56 "How can engineering be approached using fundamental scientif" ideal=20260411-elon-musk-biography.md llm→20260411-elon-musk-biography.md hybrid→20260411-leo-harness-engineering-3d.md
  • Q57 "What is the Idiot Index and how is it calculated for rockets" ideal=20260411-elon-musk-biography.md llm→20260411-elon-musk-biography.md hybrid→20260411-elon-musk-biography.md
  • Q59 "I'm not looking for general biographies but specific methods" ideal=20260411-elon-musk-biography.md llm→20260411-elon-musk-biography.md hybrid→20260411-leo-harness-engineering-3d.md
  • Q60 "Can you explain innovative approaches to production efficien" ideal=20260411-elon-musk-biography.md llm→20260411-elon-musk-biography.md hybrid→20260411-1803-03635-lottery-ticket-hypothesis.md
  • Q65 "I'm not looking for general psychology but specifically dual" ideal=20260411-thinking-fast-and-slow.md llm→20260411-thinking-fast-and-slow.md hybrid→20260411-2307-11760-emotionprompt.md
  • Q67 "How can attention mechanisms handle extremely large token co" ideal=20260411-msa-memory-sparse-attention.md llm→20260411--1706-03762-attention-is-all-you-need.md hybrid→20260411-msa-memory-sparse-attention.md
  • Q69 "Can a 4B parameter model outperform a 235B model in question" ideal=20260411-msa-memory-sparse-attention.md llm→20260411-1803-03635-lottery-ticket-hypothesis.md hybrid→20260411--1706-03762-attention-is-all-you-need.md
  • Q77 "I'm not looking for basic AI storage info, but how memory sh" ideal=20260411-hermes-evensong-synthesis.md llm→20260411-evensong-paper-zh.md hybrid→20260411-hermes-evensong-synthesis.md
  • Q78 "What are the latest findings on how AI processes memory and " ideal=20260411-hermes-evensong-synthesis.md llm→20260411-2604-08256-hypermem.md hybrid→20260411-evensong-paper-zh.md
  • Q81 "Can you find information on the Parity Principle in AI cogni" ideal=20260411-philosophy-supervisor-evensong.md llm→20260411-thinking-fast-and-slow.md hybrid→20260411-evensong-paper-zh.md
  • Q84 "What are the latest theories connecting philosophy and artif" ideal=20260411-philosophy-supervisor-evensong.md llm→20260411-philosophy-supervisor-evensong.md hybrid→20260411-philosophy-supervisor-evensong.md
  • Q87 "What is the parity principle in cognitive science discussion" ideal=20260411-clark-chalmers-extended-mind.md llm→20260411-thinking-fast-and-slow.md hybrid→20260411-clark-chalmers-extended-mind.md
  • Q88 "外部工具如何成为认知的一部分?" ideal=20260411-clark-chalmers-extended-mind.md llm→20260411-2310-08560-memgpt.md hybrid→20260411-clark-chalmers-extended-mind.md
  • Q89 "I'm not looking for internal memory theories, but for ideas " ideal=20260411-clark-chalmers-extended-mind.md llm→20260411-2604-08256-hypermem.md hybrid→20260411-2310-08560-memgpt.md
  • Q90 "Can objects outside us influence how we think and believe?" ideal=20260411-clark-chalmers-extended-mind.md llm→20260411-evensong-paper-zh.md hybrid→20260411-clark-chalmers-extended-mind.md
  • Q92 "What are effective methods to critically analyze foundationa" ideal=20260411-evensong-first-principles-audit.md llm→20260411-elon-musk-biography.md hybrid→20260411-evensong-first-principles-audit.md
  • Q93 "Can you explain the impact of cultural differences on stress" ideal=20260411-evensong-first-principles-audit.md llm→20260411-elon-musk-biography.md hybrid→20260411-thinking-fast-and-slow.md
  • Q96 "What are some strategies for identifying flaws in AI memory " ideal=20260411-evensong-first-principles-audit.md llm→20260411-2604-08256-hypermem.md hybrid→20260411-competitor-landscape.md
  • Q99 "What is the SSOT routing table in AI agent engineering?" ideal=20260411-leo-harness-engineering-3d.md llm→20260411-2310-08560-memgpt.md hybrid→20260411-msa-memory-sparse-attention.md
  • Q101 "I'm not looking for basic AI tutorials but specific scaling " ideal=20260411-leo-harness-engineering-3d.md llm→20260411-leo-harness-engineering-3d.md hybrid→20260411-leo-harness-engineering-3d.md
  • Q104 "How does memory influence the actions of artificial intellig" ideal=20260411-competitor-landscape.md llm→20260411-2604-08256-hypermem.md hybrid→20260411-2310-08560-memgpt.md
  • Q106 "智能代理的记忆如何影响其行为表现?" ideal=20260411-competitor-landscape.md llm→20260411-evensong-paper-zh.md hybrid→20260411-evensong-paper-zh.md
  • Q108 "Can you show me the latest research on AI memory systems and" ideal=20260411-competitor-landscape.md llm→20260411-2604-08256-hypermem.md hybrid→20260411-competitor-landscape.md