Corrections and Editorial Notes
Corrections by record: what changed, and why.
How to read corrected records
In a corrected entry or record, a bracketed marker after a passage names the change behind it: [Cnn] a correction (listed below and in the record's correction register), [E01] an editorial addition adopted by the editor, and [ST01] a structural rearrangement into the AES section order. Unmarked text is the original's own wording. In the lexicon records, the published short text is reproduced unchanged and a corrected record line gives the canonical wording where a correction applies. The historical wording of every original is in the Historical Lexicon Archive.
AIK-2025-F-P-COTPROMPT
- Corrections: AIK-2025-F-P-COTPROMPT/C01 FE — "Post-2023" heading becomes "2023 onward" · /C02 CD — Chain-of-Draft reference completed · /C03 CF — Trace-of-Thought reference identified (ACL 2024 SRW) · /C04 UC — Trace-of-Thought description separated from training-based distillation ("evaluated on the GSM8K and MATH benchmarks") · /C05 CD — sources supplied for the "related technical reports" · /C06 UC — diminishing returns scoped; smaller-models claim restated · /C07 UC — improved-performance claim scoped (MMLU, commonsense) · /C08 UC — interpretability claim qualified with the faithfulness caveat · /C09 FE — zero-shot CoT is not "zero cost" · /C10 CD — full author lists, venues and URLs · /C11 PR — code repositories for ToT and GoT supplied · /C12 CD — Active Prompting description and reference added (ACL 2024 Long Papers citation) · /C13 FE — ToT planning-and-search claim checked; no error; source attached · /C14 ST — entry restructured into AES sections · /C15 PR — boilerplate banner and closing line not carried · /C16 ED — See Also: general concepts separated from collection cross-references · /C17 ED (restoration) — version-2 passage restored in §6, aligned to Kojima et al.'s hedge · /C18 UC — "a standard technique in" becomes "used in" (editor's ruling). E01 (Table 1, §5) is an editorial addition, not a correction, and takes no correction ID
AIK-2025-F-P-MOE
- Correction IDs:
- AIK-2025-F-P-MOE/C01 FE — §5 (the original's §6.5) Mixtral row corrected to the paper's figures (MMLU 70.6%; GSM8K 74.4%, 8-shot maj@8); reference [7] extended with Jiang et al., arXiv:2401.04088
- /C02 UC — Switch Transformer row's MMLU/GSM8K figures removed; the Switch paper reports neither benchmark
- /C03 FE — Qwen3 row corrected to the Technical Report's Table 3 base-model figures (MMLU 87.81%; GSM8K 94.39%) with their evaluation conditions
- /C04 FE/CD — the ST-MoE / Expert Choice Routing conflation undone (§2, §4, §8); Zoph et al. added as reference [10]
- /C05 CD (minor) — Qwen3 release and report dates corrected (model released late April 2025; report arXiv:2505.09388, May 14, 2025)
- /C06 PR — the published formulas, restored from the formula images' alt texts in the stored post (§1, infobox, §3)
- /C07 PV — the published "Editorial Status" block is not carried into the canonical text; status is recorded as separate facts in the metadata
- /C08 FE — the generalized "≤ 10% of parameters active per token" claim (§3, §6) replaced by the range the entry's own §4 figures support (under 1% to ~28%)
- /C09 CD (minor) — §4's estimated per-token figures (†, ‡) removed; the notes state the papers do not give them
- /ST01 ST — the entry restructured into the AES eight-section order; the original's §6.5 benchmarks stand in §5
AIK-2025-F-P-CONSTAI
- Correction IDs:
- AIK-2025-F-P-CONSTAI/C01 CD — reference [3] year corrected, 2023 → 2024 ("Claude's Character," June 8, 2024)
- /C02 CD — the Omdia reference removed and both hybrid-alignment claims re-sourced to the Claude 4 System Card [7]; on the independent reviewer's scope check, the §4 Applications sentence is narrowed to Claude 4 (the Claude Sonnet 4.5 System Card documents RLHF and RL from AI feedback and never mentions Constitutional AI — verified on the official PDF) and the "most capable models" sentence is attributed to Anthropic
- /C03 UC — the unattributed forecasts (day-long autonomous work by mid-2026; Nobel-level contributions by 2027) removed
- /C04 CD — the Llama 3-8B claim cited to Zhang, arXiv:2504.04918, added as reference [8]
- /C05 CD — the MongoDB claim cited, added as reference [9]
- /C06 CD (minor) — both principles printed inside quotation marks replaced with Bai et al.'s verbatim RL principles (Appendix C)
- /ST01 ST — the entry restructured into the AES eight-section order (§5, Evidence and Evaluation: not in original). E01 is an editorial addition, not a correction, and takes no correction ID: the balance qualifier appended to the Llama 3-8B passage ("though in that study the gain came with a 9.8% drop in helpfulness and signs of model collapse in the final model"), adopted by the editor
AIK-2025-L-P-LEX2025N3
- Item-level correction IDs (each attaches to its occurrence record): AIK-2025-S-P-SPECDECODE/C01 FA — Speculative Decoding's appearance corrected to November 30, 2022 · AIK-2025-S-P-COVE/C02 FA — Chain-of-Verification's appearance corrected to 2023, and the headword's abbreviation corrected to the paper's "CoVe" · AIK-2025-S-P-MADEBATE/C03 FA — Multi-Agent Debate's appearance corrected to May 2023 · AIK-2025-E-P-COMPOUNDAI/C04 FA/CD (minor) — Compound AI's appearance corrected to February 18, 2024, with point-of-use attribution to BAIR (Zaharia et al.) at the definition, the worked example, and the Databricks statistics · AIK-2025-E-P-MECHINTERP/C05 FE — the TransformerLens "version 3 released in September 2025" clause removed (3.0 alpha pre-releases were public in 2025; no stable v3 existed at publication; stable v3.0.0 is April 17, 2026) · AIK-2025-E-P-TTCSCALE/C06 PR — the Test-Time Compute entry's Mathematical Framework equations, restored from the formula images' alt texts in the stored post · AIK-2025-E-P-TTCSCALE/C07 UC — Test-Time Compute Scaling: the "Applications" AIME 2024 figure qualified with the selection procedure that produced it (OpenAI, "Learning to reason with LLMs," September 12, 2024: 93% after reranking 1,000 samples per problem with a learned scoring function; 74% for a single sample, 83% for consensus among 64 samples); the top-500 comparison is retained against the 93% score
AIK-2025-F-P-CALM
- Corrections: AIK-2025-F-P-CALM/C01 CD — Leviathan et al. reference corrected (full authors; true title "Fast Inference from Transformers via Speculative Decoding"; ICML 2023, PMLR 202) · /C02 CD (minor) — Bowman venue corrected from "SIGNLL" to CoNLL 2016 (SIGNLL is the sponsoring SIG, not the venue) · /C03 CF — "Tsinghua AI Lab Notes Vol. 7" removed as unidentifiable (the metadata block is not carried) · /C04 UC — "Early replication studies" sentence removed (unsourced, six days after the preprint) · /C05 UC — model-scale and performance tables rebuilt from the paper's Table 1 and §7.1 (separate Transformer/CALM parameter columns; invented CALM-S row removed; the 44%/34% comparison attributed) · /C06 UC — GPU-speedup and robustness sentence removed (unsourced) · /ST01 ST — restructured into the AES eight sections. E01 is an editorial addition, not a correction, and takes no correction ID; it distinguishes this CALM (Continuous Autoregressive Language Models) from the earlier, different method "Confident Adaptive Language Modeling" (Schuster et al., 2022), adopted by the editor
AIK-2025-F-P-PDDISAGG
- Corrections: AIK-2025-F-P-PDDISAGG/C01 FA — "first formalized in DistServe" corrected: first proposed in Splitwise (2023), formalized as goodput-optimized disaggregated serving in DistServe (2024); timeline row and reference added · /C02 UC — utilization figures cited to InfoQ (prefill 90–95%, decode 20–40%); the LLaMA-70B/A800 qualifier dropped as unsourced · /C03 UC — FlowKV: the 96% figure scoped to KV-cache transmission latency (0.944 s → 0.053 s); the compression claim removed (the mechanism is pipelined transfer and contiguous allocation) · /C04 FE — "Inference Throughput Latency" corrected to Inter-Token Latency · /C05 UC — savings sourced: Splitwise's 1.4× throughput at 20% lower cost (primary); the 15–40% range kept, attributed to secondary analyses · /C06 FE — timeline rows put in date order (March Dynamo before April FlowKV) · /C07 CD — Mooncake reference title completed (main clause restored; FAST '25) · /C08 PR/PV — figure credit recorded (attachment and Copilot file name); the kimi.com link not carried (Editorial Note 3) · /C09 UC — Mooncake description corrected (disaggregated KVCache-centric store with a global scheduler, not co-location with decode clusters) · /ST01 ST — restructured into the AES eight sections (five sentences/bullets moved and marked). C10 is withdrawn; the number is retired and is not reused (the preparer's version-attribution error: the v1 and OSDI figures inverted; the original A-024 wording restored)
AIK-2025-F-P-GENUI
- Corrections: AIK-2025-F-P-GENUI/C01 FA — origin chronology corrected (Vercel September 2023 launch; "Generative UI" named October 11, 2023; AI SDK 3.0, March 1, 2024; Chen et al., August 2025; Google's November 18, 2025 implementation) · /C02 CF — the Google Research post's title, date, and content corrected (the Static/Declarative/Open-Ended taxonomy is not in the post) · /C03 CD — the Google post's four authors completed, including Vishnu Natchu · /C04 UC/CD — the Chen et al. annotation corrected and reference [8] identified as Alabduljabbar (PeerJ, October 25, 2024; 11,549 app-store reviews), not Weisz et al. (CHI 2024, a different paper) · /C05 UC — the unsupported claims removed or qualified ("third-generation Foundation Agents"; declarative generation as the most common production approach) · /C06 CD — in-text citation markers added following the post's own annotations of what each reference supports (the annotations are the contributor's claims, not independent verification — Editorial Note 4) · /C07 ED — the unmerged suggested revised lead and reference [1] excluded from the canonical text (the lead carries the C01 error; its reference omits an author) · /C08 CD — the CopilotKit reference corrected to Berman, Eli (September 12, 2025), exact title and URL (the reviewer's new finding) · /C09 CD (minor) — the Smith subtitle added ("+ how to fix it") · /ST01 ST — the entry restructured into the AES eight-section order (§4 from the post's preliminary answer; §5 and §8 marked "Not in original")
AIK-2025-F-P-ONPOLDISTIL
- Corrections: AIK-2025-F-P-ONPOLDISTIL/C01 UC — "far fewer FLOPs than RLHF" corrected to "far less compute than RL or extended supervised fine-tuning" (the post compares with SFT and RL; "RLHF" does not appear in it) · /C02 FA — §2 Origin added: GKD (June 23, 2023) and MiniLLM (June 14, 2023) as antecedents; references [P1] and [P2] added · /C03 PR — unrendered LaTeX typeset (no mathematical change) · /C04 UC/FE — the 8B/32B result corrected: benchmark (AIME'24), 150 steps, and the teacher qualification (Qwen3-8B used, compute counted as for the 32B model) stated; the RL comparison attributed to the Qwen3 replication; the full form stands (the October 27, 2025 Wayback snapshot contains the teacher sentence) · /C05 CD (minor) — the Lu et al. reference dated (October 27, 2025) · /C06 ED — See Also cross-references added (COTPROMPT, RLVR) · /C07 UC — "optionally using importance weighting or adaptive mixing with off-policy data" deleted (neither is described in [R1]) · /ST01 ST — restructured into the AES eight sections
AIK-2025-L-P-LEX2025N2
- Corrections: /C01 CD/FE — the MCP reference corrected: the fabricated "AI Standards Foundation" replaced with Anthropic's November 25, 2024 announcement; the specification's date versioning noted · /C02 CD — the Alignment Faking reference corrected to Greenblatt et al. (2024), arXiv:2412.14093 (Hubinger is the last of 20 authors) · /C03 CD/FE — the GAIA-2 reference corrected (Russell et al., Wayve, arXiv:2503.20523; "Generative AI for Autonomy" gloss noted) · /C04 CF — five unattested references removed, their items marked unsourced and kept (editor's ruling) · /C05 UC — Alignment Faking: the definition narrowed to Greenblatt et al.'s; the "major labs" claim replaced by the single dated demonstration; the RLHF/RLAIF sentence replaced by the paper's reported results; the "Inoculation Prompting" bullet removed · /C06 CD/UC — KV-cache sharing: the Apple report cited by title and identifier (arXiv:2507.13575 §2.1; within-model sharing, ~37.5% time-to-first-token reduction); the Qualcomm half removed · /C07 CD — LMGenDrive removed (no 2025 work found; the 2026 work postdates the installment); SafeKV's author corrected to Chu et al. · /C08 PR — LaTeX typeset once each; E12's fields ordered (the quotation keeps the published order) · /C09 FE (beyond the findings) — the reversed "when unmonitored" parenthetical removed (the model complied more when it believed it was in training) · /C10 UC — the Alignment Faking generalization fixed in the reviewer's own wording, with a qualifier marking the Technical Description as a simplified model · /C11 UC — the World Model "2025 breakthrough" integration claim replaced (generative world models advanced in 2025, GAIA-2 as the example; LLM integration not established) · /C12 UC — the KV-cache memory-reduction formula omitted (no derivation; its terms do not yield a memory quantity); C08's typesetting of it lapses
AIK-2025-F-P-MOD
- Corrections: AIK-2025-F-P-MOD/C01 PR — citation chips replaced with reference markers [R1] (arXiv:2404.02258) and [S1]/[S2] (clean URLs; tracking parameters removed) · /C02 CD — SkipNet's venue corrected to ECCV 2018 (was CVPR 2018) · /C03 UC — the gradient-mechanism hedge replaced with the paper's account (router weights on the gradient path, §3.4) · /C04 ED — the editor-instruction venue note not carried (venue status in Editorial Note 3) · /C05 PR — "MiD" corrected to "MoD" · /C06 PR — the U+2010 hyphen in the reference title corrected · /C07 ED — See Also cross-reference added (MOE) · /C08 CD (from the independent review) — the dead [S1] Graphcore link removed; the Router Network bullet cites [R1, §3.4]; [S2] stays as the secondary source · /ST01 ST — restructured into the AES eight sections (three sentences split from their paragraphs and marked). E01 is an editorial addition, not a correction, and takes no correction ID; it adds the paper's sampling-time routing account (§§3.5, 4.2), adopted by the editor
AIK-2025-L-P-LEX2025LAUNCH
- Corrections: /C01 FA/CD — RLAIF dated from December 15, 2022 (Bai et al., Constitutional AI); the Lee et al. reference completed (arXiv:2309.00267; ICML 2024); the "Why 2025" note records 2025 variants, not origin · /C02 CD — the Paglieri et al. authors completed (arXiv:2509.03581); Date of Appearance September 3, 2025 · /C03 CF — the Chopra et al. authors and venue corrected (AAMAS 2025; arXiv:2409.10568); no institution given as author; Date of Appearance September 14, 2024 · /C04 CF — the AgentScen paper corrected (Chen et al., AgentScen 2025 proceedings; "AgentScen" names the workshop; the headword kept); the Toronto/NRC affiliations confirmed · /C05 UC — three overclaims replaced by what their sources report (M2 MoE upcycling; M6 WebArena-Lite/WebVoyager figures; M7 Minecraft 105% improvement and more than 6× fewer ineffective actions, wording amended) · /C06 CD — two misattributed references removed, their items marked unsourced (the cited papers concern other items) · /C07 CD — the "(Medium)" RAG chip removed (no source) · /C08 FA — the scalable-oversight history rewritten with the dated lineage (Amodei et al. 2016 through CAI, December 2022); the "first widely adopted" sentence removed; Dates of Appearance June 2016 · /C09 UC — two unidentified "2025 developments" bullets removed (the planner–executor sentence kept) · /C10 PR — formulae typeset once each; chips given as clean references; tracking parameters removed (the "As published" quotations keep the chips). Review resolutions: C01 and C08 (flagged conflicts) were checked independently by the reviewer; C04's affiliation confirmed; C05's M7 wording amended; C08's Saunders et al. confirmed; the O1 chips identified (S1 Mu & Lin's MoE survey; S4 Sharma's RAG survey), S2's identity unconfirmed with the link kept. /C11 FE/CD — Theoretical Learning Dynamics of MoE: the published 'first provable results on MoE cluster learning' claim qualified to 'among the first analyses', with earlier theory noted (Chen et al., NeurIPS 2022); Kawata et al.'s full title restored. Beyond the audited findings; reviewed by DeepSeek-V4.1 · /C12 FE/UC/CD — LExI: the definition corrected (the per-layer allocation is fixed after training, from the model's weights; it does not vary per input at inference); "first adaptive-expert study" replaced with the 2025 contribution and the earlier dynamic-routing work (Huang et al., 2024); the reference's authors completed. Reviewed by Grok 4.5
AIK-2025-L-P-LEX2025N5
- Correction IDs: /C01 FE — the Gated Attention paper's award class corrected from "Runner-Up Paper" to "Best Paper" at three sites (table row, both cells; revised full entry), with the Qiu et al. citation (arXiv:2505.06708) · /C02 PV/CD — the ICML 2025 Ramdas tutorial citation restored in the Game-Theoretic Statistics full entry and the D10 table record (Ramdas, "Game-theoretic Statistics and Sequential Anytime-Valid Inference," July 14, 2025; ICML tutorial 40002) · /C03 CD — the non-citation label "ICML 2025 Attendee Reflection" replaced with the real papers for D3 (Ghosal, Maini & Raghunathan, arXiv:2507.09937, ICML 2025) and D8 (Geng et al., arXiv:2507.06187, COLM 2025). No structural correction (lexicon records in the L1/L3 format); no presentation correction. Annotations (no correction ID): (the table's D2 Best Paper claim correct; the first review's doubt wrong); (figures verified); the D8 "discussed at ICML 2025" note (unverified; the paper's venue is COLM 2025); the Hivemind cited-title note (the paper's full title ends "(and Beyond)"); D10's headword rename and the elided drafts
AIK-2025-L-P-LEX2025N7
- Correction IDs: /C01 FE/UC — the "300%" Market Impact bullet removed; the multi-agent timeline corrected to 2023 (AutoGPT released March 30, 2023; CrewAI repository created October 27, 2023), the bullet moved before the 2024 Devin note · /C02 CD/FE/UC ⚑ — the "7–12 tools daily" sentence removed and its successor reworded to stand alone; the LangChain/LlamaIndex specification bullet removed; the invented "AI Interoperability Consortium" citation replaced with Anthropic's November 25, 2024 announcement and the date-stamped specification versions · /C03 FE/CF — AlphaProof re-dated to 2024 (announced July 25, 2024; with AlphaGeometry 2, silver-medal standard at IMO 2024); "GitHub's Copilot Reasoning" removed · /C04 CD (minor) — Rahimi reclassified as a Pythonic AI company-blog post; Venkiteela as a Computer Fraud & Security journal article (2025(1), 1259–1271; DOI 10.52710/cfs.817) · /C05 PR — the three bracketed-TeX Formal Expressions typeset as display formulas, content preserved. No structural correction (lexicon records in the L1/L3 format). Annotations (no correction ID): (the published full texts are Copilot's, not Kimi's announced set; no Constitutional AI full text exists); (the list's "no repetition" claim fails at 9 of 12); (the first review's "pioneered in 2023" error, inside an Accuracy 10/10 rating); the byline's "ChapGPT" (frame text, not carried; ChatGPT spelled correctly in the records)
AIK-2025-L-P-LEX2025N4
- Correction IDs: /C01 FE/FA — Vibe Coding coinage stated (Karpathy, X, February 2, 2025; date confirmed from the post identifier; wording as quoted by secondary sources) in both states; Date of Appearance corrected · /C02 CD/FE — Gartner forecast restored (by 2026, 30% of enterprises will automate more than half of their network activities; press release, September 18, 2024) in both states; citation title restored in full · /C03 UC/FE-minor — VeriTrail removed as an enterprise GraphRAG implementation in both states (Microsoft Research hallucination-detection method, arXiv:2505.21786) · /C04 UC — three unattributed figures removed without replacement (G3 50% employee involvement; G10 $11.8 billion in both states; G12 51% of users) · /C05 UC narrowed — Samsung example kept generic, citing AI Ecobubble (Samsung Newsroom India, March 13, 2024); the "basic automation" characterization recorded as the entry's assessment · /C06 UC (review observation §5) — unattributed 62% figure removed (G6 Text 2) · /C07 UC — AI Washing (G2): the "Why It Matters" adoption figure restated with the surveyed population and threshold (McKinsey, The State of AI, 2025: 88% of 1,993 respondents, surveyed June 25–July 29, 2025, reporting regular AI use by their organizations in at least one business function); the figure measures organizational AI use, not AI washing · /C08 UC — Shadow AI (G3): the "Why It Matters" firm-share figure restated with the surveyed population and threshold (McKinsey, The State of AI, 2025: 88% of 1,993 respondents, surveyed June 25–July 29, 2025, reporting regular AI use by their organizations in at least one business function); the figure measures organizational AI use, not shadow AI. No structural correction (lexicon records in the L1/L3 format). Annotations (no correction ID): (prominence framing; pre-2025 subjects under the "took off in 2025" premise); (Shadow AI 2026-entry link, verified — post 16017, October 2, 2026); Review Applied? = Not applied; the bar question ruled for G1 (established); unstated release numbers recorded
AIK-2025-F-P-CIRCUITSP
- Correction IDs:
- AIK-2025-F-P-CIRCUITSP/C01 — conditional computation distinguished from circuit sparsity (three sites: taxonomy, Gemini 1.5 bullet, Open Questions bullet)
- /C02 — Gao et al. reference completed (arXiv:2511.13653)
- /C03 — Limitations claim completed to "training and inference" compute per Gao et al. §5, with attribution
- /ST01 ST — the entry restructured into the AES eight-section order (§§5/6/8 marked "Not in original"). No editorial addition. The "defining source" parenthetical was removed.
- Blind-audit annotations: thirteen annotations from GLM 5.3 Flash's blind independent audit, ruled on and approved by the editor and entered in the correction register (twelve new annotations, plus a corroboration sentence on the existing Mathematical framework annotation). No new correction ID; C01–C03 and ST01 stand; the published texts are unchanged.
New additions (origin N) carry review histories, not correction IDs; their dated records are in Appendix G.