Research.

Tao An. Poster at the NeurIPS 2026 Workshop on AI for Meta-Science (AI4MetaScience; non-archival). A retrieval-grounded, protocol-frozen novelty audit of 166 machine-generated papers (FARS) against 166 topic-matched ICLR 2025 submissions. Each paper is decomposed into contribution claims on four facets (purpose, mechanism, evaluation, domain; 549 machine and 494 human contributions); prior art is retrieved under per-paper submission-date cutoffs; contributions are classified as covered, recombination, or facet-novel under a pre-registered two-judge protocol. Machine contributions are judged facet-novel more often than human ones (56.1% vs. 35.6%), and the gap rests mainly on the purpose facet: requiring an uncovered facet other than purpose shrinks it from 20.5 to 4.1 points (95% CI −0.3 to 8.5). The judge layer cannot be certified by another LLM: adversarial re-auditors return refutation rates from 0% to 100% on the same items depending on auditor model and prompt. A 106-pair gold prior-art audit finds the deployed retrieval surfaces the known prior art for only 25–29% of pairs per arm, so automated novelty rates are exploratory upper bounds. A companion integrity audit of 306 Agents4Science 2025 submissions finds hard fabrication evidence in 0/47 accepted versus 16/197 rejected (one-sided Fisher p = 0.029; 0.072 under conservative coding), an association that reflects both AI-reviewer detection and authors' own disclosure.

Tao An. Transactions on Machine Learning Research (TMLR), 2026. Surveys intervention timing for always-on assistants around one decision rule: intervene iff the expected benefit of acting exceeds the expected cost of interrupting. Reconnects two literatures that do not cite each other, the 1999–2017 interruptibility line that formalized interruption cost but had no capable actor, and the 2024–2026 proactive-agent wave that has actors but rediscovers the cost term only in fragments. Argues evaluation is the gating layer, and proposes a benchmark design for open-world intervention timing with an explicit cost term.

Tao An, Shuai Feng. Transactions on Machine Learning Research (TMLR), 2026. Sampling an aligned LLM 25 times per prompt and embedding the completions, output dispersion (mean pairwise cosine distance) is predicted by the consensus density of the task, scored by held-out judges that see only the prompt (Spearman ρ = −0.91). The task-type ordering holds on two unrelated models (ρ = 0.85 and 0.84), the judge relation holds within two professional domains (−0.83 over 42 pharmaceutical prompts, −0.62 over 40 software-engineering prompts), and a preliminary expert panel tracks the judges at ρ = 0.74. A matched base-vs-instruct comparison shows alignment amplifies a gradient the pretrained base already carries, compressing the high-consensus end ~8.7-fold against 1.8-fold at the low end. The link to AI equalizing or amplifying human outcomes is stated separately, as an untested hypothesis.

Tao An. Published in the proceedings of the 5th International Conference on Hybrid Human-Artificial Intelligence (HHAI 2026, Brussels), IOS Press Frontiers in Artificial Intelligence and Applications vol. 423, pp. 212–220 (open access, CC BY-NC). Drawing on structured field observations since mid-2024, this position paper reconciles the “AI as equalizer” and “AI as amplifier” debates: AI narrows novice–expert gaps on routine, well-structured tasks but amplifies them on complex tasks requiring deep judgment. Domain expertise, not prompt engineering, determines who benefits most from AI.

Tao An. Links 20,529 arXiv papers published 2015–2025 to their author-designated GitHub repositories and estimates the rank association between annual fork flow and annual citation flow in every cohort-by-period cell. The association falls from roughly 0.45–0.50 in the late 2010s to roughly 0.25 by 2024, and the variation sits on the calendar-period axis rather than the publication-cohort axis: with period included the cohort coefficient is −0.0017 (p = 0.74), and the 2015 cohort, the same 262 papers throughout, declines from 0.49 to 0.04 over its own lifetime. Two candidate explanations fail on the correct axis. There is no discontinuity at the public release of general-purpose language models (−0.002, 95% CI [−0.061, +0.057]) and no gradient in field-level exposure to them (p = 0.82); nor does the citation mix degrade, with the share of substantively influential citations flat across the decade. Enumerating the full frame of 160,150 papers rather than sampling from it, the association among works that are both substantially reused and substantially cited shows no trend (−0.0031 per year, 95% CI [−0.0156, +0.0094], n = 3,885), an interval that excludes the full-sample estimate: the decoupling is a property of the population, not of its high-impact tail.

Tao An. A controlled ablation isolating the stored memory representation inside one fixed retrieve-rerank-reason pipeline: LLM-extracted typed artifacts versus verbatim conversation chunks, holding the model, retriever, reranker, and judge constant. Verbatim chunks win by 15.9 points on LoCoMo (43.9% vs. 28.0%) and 22.0 points on LongMemEval-S (67.4% vs. 45.4%); the extracted-artifact pipeline never beats naive RAG. The mechanism is lossy distillation: extraction discards verbatim detail that chunks retain for free, so structured memory should augment verbatim text, not replace it. (Formerly titled It's Fidelity, Not Structure.)

Tao An. Proposes Cognitive Workspace, an alternative to passive RAG modeled on working memory: active memory management, hierarchical cognitive buffers, and task-driven context optimization. The prototype reuses its own curated state on 58.6% of turns (vs. 0% for RAG), a 17–18% net efficiency gain. It did not test answer accuracy against a verbatim baseline; the 2026 Fidelity Before Structure ablation later found that write-time curation loses to verbatim chunks on accuracy.

Tao An. Audits both promises a citation makes on the ACL 2026 proceedings treated as a census: 4,459 papers and 209,985 references. Existence holds: 91.0% of references resolve, and two references (0.001%) are confirmed nonexistent. Claim support, scored by a two-stage language-model judge, does not repeat. An audit of 2,110 claim citations from 100 papers put the confirmed support-defect rate at 0.95%; two further independent draws at the same commit returned 5.66% and 6.12%, pooling to 5.90% [5.12, 6.80] over 3,033 claim citations (7.9 SE from the audited figure). The gap sits in the support-judgment layer and traces to a first-pass judge whose model was read from an unlogged environment variable. The paper reports all three runs and takes the non-replication, not any single rate, as the principal finding: pin and log the judge, and repeatability returns. The existence census is unaffected.

Tao An. An independent replication and reframing of Anthropic's “global workspace” (Jacobian-lens) claim on open Qwen3 models. A mouth-exclusion audit, which scores a lens readout as covert only when its token is far outside the model's own next-token distribution, splits the workspace: covert content survives almost only for context registers (the conversation's language, a corrupted word's intended form), while content plans (rhyme, arithmetic, associations) fall to a permutation floor. The surviving registers are causally load-bearing under amplitude-matched steering, and workspace edits rewrite the model's representation of the question itself across a 1.7B–14B scale ladder. The register and capture findings reproduce on a second architecture (Gemma-2-2B).

Tao An. Defines GEO-poisoning detection and attribution for Chinese generative search (DeepSeek, Doubao, Kimi): a five-technique taxonomy of coordinated, inauthentic manipulation, a task reframing from attack-success to detection → classification → account-cluster attribution, and a legally-constructed synthetic benchmark. A provenance pilot shows detection and attribution need different substrates: content features detect that manipulation happened (F1 0.93), but only an account-interaction graph attributes it to a seller cluster (0.96). A confidence-gated fusion covers the taxonomy where a learned GNN and a zero-shot LLM both fail.

Tao An. Chinese invention patent, under examination. GNN-based recommendation over heterogeneous data–information graphs.