docs: sync BENCHMARK.md + README to paper (arXiv 2605.13438v3) - #29
Merged
Conversation
The benchmark docs had drifted badly from the technical report — most notably LongMemEval was still listed as "0% (retrieval returns empty context)" when the paper reports 93.0% J-Score overall (N=500). Align all headline numbers to the paper as the source of truth: - LongMemEval: 0% -> 93.0% overall, with per-category J-Scores and the build/answer/judge stack; cross-ref PR #26/#27 for the MS lever. - LoCoMo: 82.8% (gpt-4.1-mini) -> 81.23% overall (paper Table 4 stack, gpt-4o-mini) vs ENGRAM / MemOS / Zep. - MuSiQue F1 -> 58.7 vs HippoRAG 2 49.3. - BABILong -> 85.0 vs ARMT (fine-tuned) 83.8. - ToMi -> 83.5 vs AutoToM 80.2. - Add a canonical "Paper results" section (Tables 3-5 + Fig. 4) to BENCHMARK.md and note that the reported stack is the proactive-substrate configuration, not a per-benchmark tuned ceiling. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HCToJStvPhZbWGY2ZcBfy7
duanyiqun
changed the title
duanyiqun
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
The benchmark docs had drifted badly from the technical report (arXiv:2605.13438v3). Most glaringly,
docs/BENCHMARK.mdstill listed LongMemEval as "0% — retrieval returns empty context", while the paper reports 93.0% J-Score overall (N=500). This PR re-syncs both the public README and the internaldocs/BENCHMARK.mdto the paper, which is treated as the source of truth.Changes
0%→ 93.0% overall, with per-category J-Scores (SSA 100.0 / SSU 97.1 / KU 94.9 / SSP 93.3 / MS 91.0 / TR 88.7) and the build/answer/judge stack. Cross-references PR Neural-symbolic MS agent (experimental, OFF by default, shelved) #26/MS: neural-symbolic take-max router + W1 writer fixes #27 for the MS count/sum lever.82.8%(gpt-4.1-mini) → 81.23% overall on the paper's Table 4 stack (gpt-4o-mini), vs ENGRAM 77.55 / MemOS 75.80 / Zep 75.14.BENCHMARK.md.Note on the lowered numbers
A few are lower than older internal numbers (ToMi 91.6% EM, BABILong 96.0% EM). That is deliberate: the reported configuration is the proactive-substrate stack, not a per-benchmark tuned ceiling. The full rationale (and the cognitive-bias framing behind it) is being proposed separately as a Design Philosophy PR for review.
🤖 Generated with Claude Code
https://claude.ai/code/session_01HCToJStvPhZbWGY2ZcBfy7