Pith. sign in

REVIEW 4 major objections 5 minor 32 references

Untuned reciprocal rank fusion outperforms three adaptive per-query weighting schemes on 10-K evidence retrieval, once retrieval units are cut to the encoder's window.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 01:03 UTC pith:ZW25QCZ4

load-bearing objection A solid, honest empirical study: the evidence-unit fairness point is real, the router null result is credible, and the label mapping is the main caveat—worth serious referee time. the 4 major comments →

arxiv 2608.00183 v1 pith:ZW25QCZ4 submitted 2026-07-31 cs.CE cs.AI

Evidence-Unit Fairness and the Limits of Query-Adaptive Sparse-Dense Fusion in Financial Document Retrieval

Chenyu Wu , You Lin This is my paper
classification cs.CE cs.AI
keywords financial document retrievalsparse-dense hybrid retrievalreciprocal rank fusion10-K filingsevidence-unit fairnessquery-adaptive fusiondense retrievalretrieval evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Retrieving answers from corporate 10-K filings is hard because queries are short and the evidence sits deep inside long, table-heavy documents. This paper establishes two things. First, comparisons between lexical and dense retrieval are unfair unless the retrieval unit is cut to the dense encoder's reading window; on the FinDER benchmark, nearly half the labeled evidence begins beyond that window. Second, once that mismatch is fixed, fusing BM25 with a compact dense encoder improves reference-level Hit@10 by roughly 28 percent over either component alone, and an untuned reciprocal rank fusion scores highest of all tested methods. Three lightweight per-query weighting schemes fail to reliably beat the fixed equal-weight blend, despite an oracle showing 21.8 percent headroom, so the paper argues simple fusion is the defensible default and the next gains lie in better evidence-unit construction.

Core claim

The paper's central claim is that the apparent performance gap between sparse and dense retrieval on financial filings is partly a measurement artifact: when the retrieval unit is a large table or section, a dense encoder with a 512-token budget never sees most of the labeled evidence, while a lexical index sees all of it. After segmenting the corpus into overlapping windows that fit the encoder budget and mapping each expert reference to its windows, BM25 and a compact dense encoder score nearly the same alone (reference-level Hit@10 around 0.139 and 0.142), and fusing them raises the score to 0.180–0.189, a relative gain of roughly 28 percent. Reciprocal rank fusion with no training and no

What carries the argument

The central mechanism is the evidence-unit fairness correction: segmenting every parsed filing unit into overlapping 160-word windows that fit the dense encoder's 512-token input, and mapping each labeled reference to its windows with provenance, so reference-level Hit@10 can be measured fairly. This correction turns a confounded comparison into one where the two retrievers are equally exposed to the evidence. The second working part is fusion: reciprocal rank fusion (combining ranked lists by summing reciprocal ranks with a fixed offset, here k=60) and weighted score fusion with an interpolation weight. The third is the oracle/router apparatus: a per-query grid oracle defines an upper bound

Load-bearing premise

The automatic reference-to-window mapping is accurate enough for the comparison; only 89.1 percent of queries resolved, with literal containment for just 30.3 percent of references and no verified precision figure, so systematically mis-mapped labels—especially in table-heavy accounting questions—could distort the measured fusion gains and the routers' failure.

What would settle it

Manually audit a random sample of the 5,082 resolved query-reference-window mappings, stratifying by query category and by the literal-containment diagnostic; if mis-mapping concentrates in table-heavy categories where BM25 and fusion appear to excel, recompute Hit@10 on verified mappings and check whether the 28 percent fusion gain and the router null results survive. A second check: rerun the router comparisons only on the 1,518 literal-containment queries with the full protocol; the paper already reports the router effects shrink there, so a confirmed null on a fully verified subset would s

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • On encoder-windowed 10-K corpora, fusing BM25 with a compact dense encoder improves reference-level Hit@10 by roughly 28 percent over either component; the equal-weight blend is within a flat optimum, so it is not sharply tuned.
  • Untuned reciprocal rank fusion is the highest-scoring deployable method in this study, exceeding the equal-weight blend by 0.009 in reference-level Hit@10 with a clustered 95% confidence interval excluding zero, and requires no training or weight selection.
  • Per-query weighting is not worth the complexity here: none of the three lightweight routers reliably beats the fixed blend, even though the grid oracle shows 21.8 percent headroom.
  • A category-based hard switch between pure retrievers is also dominated by the fixed blend (0.158 versus 0.180), so the most direct use of category information does not beat simple fusion.
  • Tuning the fixed weight on a small validation split can slightly underperform simply setting it to 0.5, reported as suggestive, reinforcing the untuned-default recommendation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the automatic label mapping is accurate, these results suggest that the main yield left on the table in financial-filing retrieval is not router sophistication but evidence-unit construction—keeping table headings attached to their numeric rows—and the same lesson should transfer to any retrieval pipeline over XBRL-sourced documents.
  • Editorial inference: the router failure is bounded by reachability: 77.6 percent of queries are unanswered under any weight, so even a perfect per-query weighting can only affect the reachable minority; a router that first scores reachability, not just weight, would be a more direct test of the headroom.
  • Editorial inference: the fairness correction is dataset-agnostic; any dense retriever used over long documents should be evaluated on encoder-sized windows, and this paper gives a template for reporting reference-level provenance that other benchmarks could adopt.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper studies sparse-dense hybrid retrieval over FinDER, a financial 10-K retrieval benchmark, and makes two main claims. First, it identifies an 'evidence-unit fairness' problem: the parsed retrieval units are much larger than the 512-token input limit of the dense encoder, so 46.2% of labeled evidence begins at or beyond word 512, biasing comparisons against dense retrieval. The authors segment the corpus into 160-word overlapping windows and automatically map references to windows. On this corrected corpus, BM25 and a compact dense encoder each achieve reference-level Hit@10 near 0.14, while fusion roughly doubles either component; reciprocal rank fusion (RRF) scores highest at 0.1892, ahead of the equal-weight blend at 0.1802. Second, the paper asks whether per-query fusion weights help: a grid oracle shows 21.8% headroom, but three lightweight routers (a score-confidence heuristic, a random forest, and a ridge regressor over embeddings) fail to beat the equal-weight blend reliably under company-grouped cross-validation with cluster-robust inference. The paper concludes that untuned RRF is a sensible default and that effort may be better spent on evidence-unit construction than on lightweight per-query weighting.

Significance. If the results hold, the paper makes a useful methodological contribution: it quantifies a real retrieval-unit confound and offers a clean protocol for removing it. The evaluation design is a strength: company-grouped cross-validation with clustered bootstrap inference, a fixed decision rule for the router comparisons, an explicit oracle upper bound, and a literal-containment sensitivity subset. The negative result on lightweight query-adaptive fusion is also valuable, especially because the equal-weight blend is simple and the routers are described precisely enough to reproduce. The paper is not circular: the fixed baselines are training-free or independently tuned, and the routers' failure is the claim rather than a fitted success. However, the central comparison rests on automatically constructed relevance labels with no verified precision and with 10.9% of queries excluded; this makes the current evidence conditional rather than definitive. A manual label audit and analysis of excluded queries would substantially increase confidence.

major comments (4)
  1. [Section III-C/III-D; Tables I-II] The automatic reference-to-window mapping is the load-bearing component of the evaluation. It resolves only 5,082 of 5,703 queries (89.1%) and achieves literal containment for only 30.3% of references; no verified precision figure is reported. The paper's own qualitative inspection says failures concentrate in financial-statement tables, which is exactly the category where BM25's exact-number matching is expected to help. If unresolved or silently mis-mapped references are systematically distributed, the measured +0.0090 RRF advantage and the null router differences could be artifacts of label construction rather than retrieval quality. The literal-containment subset (n=1,518) is a reasonable robustness check for the fusion-vs-components comparison, but it cannot speak to the 621 excluded queries or to references that are mapped incorrectly despite high overlap. I ask for a manual audit
  2. [Section V-D; Table II] The only comparison that is statistically reliable on the full set — RRF over the equal-weight blend — is not reliable on the cleaner literal-containment subset. The paper reports a similar point estimate (+0.0092) but p=0.11 on that subset. Given that the practical guidance in Section VI recommends RRF as the default, the evidence for that recommendation is weaker than the full-table result implies. The revision should either temper the practical claim or provide additional evidence, such as a larger manually verified label set or an analysis showing that the full-set advantage is not driven by label noise.
  3. [Section IV-B; Table II] The router comparison is statistically careful, but the study appears underpowered to detect small positive gains. The ridge router's one-sided p is 0.0555, close to the unadjusted 0.05 level, and the random-forest estimate is effectively zero with a wide interval. The conclusion that no router 'establishes' an improvement is formally correct, but the practical advice 'do not assume routing will help' goes further. I request a minimum-detectable-effect analysis under the company-clustered bootstrap, or a pre-specified equivalence/non-inferiority margin, so readers can distinguish 'no evidence of benefit' from 'evidence of no benefit.'
  4. [Section III-D; Section V-D] The paper states that mapping failures concentrate in financial-statement tables and that the heading-only mapping may make some numeric queries' answer-bearing rows absent from any retrievable window. This is an important caveat, but it is not quantified for the 621 excluded queries. If the excluded queries are disproportionately numeric, the evaluation set is biased toward narrative queries, which would directly affect both the fusion gain and the router nulls. Please report the category distribution and other query features for the resolved versus unresolved sets, and if possible bound the impact of the exclusions on the main comparisons.
minor comments (5)
  1. [Section III-B] The phrase 'these audited pairs' is unclear; define what 'audited' means here and how the 46.2%/55.2% figures were computed.
  2. [Section IV-B] The training target description contains a likely typo: 'we take the best rank ... and computer d/(rb + rd)' should probably be 'compute' with the formula written consistently. Also clarify whether the exclusion of queries where neither retriever finds a relevant window in the top 100 applies only to the training set or also to evaluation.
  3. [Table I] The ridge router row reports 'n/a' for window-level Recall@10 without explanation. Since the other rows report this secondary metric, please either compute it or state why it is not applicable.
  4. [Figure 1 caption] The explanation of the α=1 endpoint difference from the BM25-only row is useful, but it would be clearer in the main text rather than only in the caption.
  5. [Section IV-C] The sentence describing the Bonferroni adjustment is convoluted: 'each fixed before that router's own final evaluation and tested in sequence rather than planned jointly at the outset.' Please clarify whether the three router tests were planned sequentially or jointly, since the stated correction assumes joint testing of three methods.

Circularity Check

0 steps flagged

No significant circularity: the retrieval comparisons are self-contained; fitted routers are evaluated against an independent fixed baseline, and the oracle is explicitly an upper bound rather than a prediction.

full rationale

Walking the derivation chain, no claimed result reduces to its own inputs. The corpus windowing and automatic reference-to-window mapping (Section III-C) are benchmark-construction steps with stated limitations (89.1% query resolution, 30.3% literal containment, no verified precision); these are label-quality caveats, i.e., potential correctness risks, not circular reductions. The fixed baselines are training-free: BM25 uses standard parameters, the dense encoder is a released model, the α=0.5 blend is untuned by design, and RRF uses its usual offset with no per-query fitting. The adaptive routers are trained on a rank-based target derived from retrievers' outputs, but the paper's central claim is that none of them beats the independent fixed blend under company-grouped cross-validation; a fitted method failing against an external baseline cannot be circular. The grid oracle is explicitly labeled 'an empirical upper bound for any router restricted to this same grid,' so its 21.8% headroom is not presented as a deployable prediction. There are no load-bearing self-citations: the reference list contains no work by the present authors, and FinDER [18] is an external benchmark. The reader's flagged concerns about label noise and the 621 excluded queries are legitimate correctness risks but are not instances of the derivation being equivalent to its inputs. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The central empirical claims rest on FinDER label quality, the automatic window mapping, the choice of e5-small-v2 as representative dense encoder, and the company-grouping/independence assumption; none of these is derived in the paper. There are no invented entities or new modeling constructs.

free parameters (7)
  • window size and overlap = 160 words, 32-word overlap
    Section III-C; chosen to fit the 512-token encoder budget; changing window geometry changes every retrieval result.
  • RRF rank offset k = 60
    Section IV-A; fixed 'usual offset' from Cormack et al.; no per-query tuning, but the value affects RRF scores.
  • equal-weight blend alpha = 0.5
    Section IV-A/B; the fixed baseline; the paper shows it sits on a flat near-optimal plateau.
  • BM25 k1 and b = k1=1.5, b=0.75
    Section IV-A; standard Lucene defaults; affects the lexical score distribution.
  • candidate top-k = 100 per retriever
    Section IV-A; fusion scores the union of both top-100 lists; a larger k could change the comparison.
  • oracle grid step = 0.05
    Section IV-B; the headroom estimate of 21.8% is an upper bound over this grid, not over all possible weights.
  • router validation hyperparameters = RF max_depth in {3,6,None}, min_samples_leaf in {5,10,20}; ridge alpha in {1,10,50,100,300,1000}, chosen by validation H
    Section IV-B; the router negative result is contingent on these selected values and on the validation procedure.
axioms (6)
  • domain assumption FinDER expert-annotated query-evidence pairs are approximately correct ground truth.
    Section III-A/D; the paper relies on released labels and only automatic diagnostics, with no human verification of the mapping.
  • domain assumption The automatic prefix-key mapping correctly locates references in windows for the 5,082 resolved queries.
    Section III-C; mapping is algorithmic and approximate; only 30.3% of references pass literal containment, so most mappings are not literally confirmed.
  • domain assumption e5-small-v2 is a representative compact dense encoder for financial retrieval.
    Section IV-A and VI-B; chosen for reproducibility on modest hardware; numeric levels may differ with other encoders.
  • domain assumption Company-level grouping makes company clusters an acceptable independence unit.
    Section IV-C; cluster-robust bootstrap treats 489 companies as independent, though same-industry firms may share boilerplate language.
  • domain assumption 160-word windows preserve enough passage context for retrieval.
    Section III-C; the correction assumes window-level relevance mapping is meaningful; windows that split table headings from rows break retrievability, as the authors note in Section III-D.
  • domain assumption BM25 tokenization and Lucene scoring are a fair lexical baseline.
    Section IV-A; a different tokenizer or stopword policy could shift the lexical baseline.

pith-pipeline@v1.3.0-alltime-deepseek · 9778 in / 12994 out tokens · 135760 ms · 2026-08-04T01:03:43.309568+00:00 · methodology

0 comments
read the original abstract

Retrieval over financial filings is difficult because queries are short and acronym-heavy while the answer-bearing evidence sits inside long, table-dense documents. We study sparse-dense hybrid retrieval on FinDER, a benchmark of expert-annotated questions over corporate 10-K filings. Our first finding is methodological: if the retrieval unit is larger than the dense encoder's input window, the dense model never sees a large share of the labeled evidence, confounding comparison against a full-text sparse baseline. We measure this directly and remove it by segmenting the corpus into encoder-sized windows. On the corrected corpus, fusing BM25 and a compact dense encoder improves reference-level Hit@10 by roughly 28 percent over either component, and training-free, untuned reciprocal rank fusion exceeds the equal-weight blend in an exploratory comparison. We then ask whether choosing the fusion weight per query helps: an oracle over the interpolation-weight grid shows headroom of 21.8 percent, yet none of the three lightweight adaptive routers (a score-confidence heuristic, a random forest over query features, and a ridge regressor over query embeddings) establishes a statistically reliable improvement over the fixed blend under company-grouped cross-validation with cluster-robust inference. Simple fusion is a strong baseline here, and we discuss why per-query weighting does not capture the available headroom.

Figures

Figures reproduced from arXiv: 2608.00183 by Chenyu Wu, You Lin.

Figure 1
Figure 1. Figure 1: Reference-level Hit@10 as a function of the fixed interpolation weight. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 1 canonical work pages

  1. [1]

    Financial distress and jump tail risk: Evidence from China’s listed companies,

    X. Liu, Y . Zhang, M. Tian, and Y . Chao, “Financial distress and jump tail risk: Evidence from China’s listed companies,”Int. Rev. Econ. Finance, vol. 85, pp. 316–336, 2023

    2023

  2. [2]

    Interpretable factor decomposition for decision intelligence in large-scale financial markets: Evidence from China’s A-share market,

    X. Han, Y . Xiao, Z. Zhang, and M. Zheng, “Interpretable factor decomposition for decision intelligence in large-scale financial markets: Evidence from China’s A-share market,” arXiv:2606.12843, 2026

    Pith/arXiv arXiv 2026

  3. [3]

    Some simple effective approximations to the 2-Poisson model for probabilistic weighted retrieval,

    S. E. Robertson and S. Walker, “Some simple effective approximations to the 2-Poisson model for probabilistic weighted retrieval,” inProc. 17th Annu. Int. ACM SIGIR Conf., 1994, pp. 232–241

    1994

  4. [4]

    The probabilistic relevance framework: BM25 and beyond,

    S. Robertson and H. Zaragoza, “The probabilistic relevance framework: BM25 and beyond,”Found. Trends Inf. Retrieval, vol. 4, no. 1–2, pp. 1–174, 2009

    2009

  5. [5]

    Dense passage retrieval for open-domain question answering,

    V . Karpukhin, B. O ˘guz, S. Min, P. Lewis, L. Wu, S. Edunov, D. Chen, and W. Yih, “Dense passage retrieval for open-domain question answering,” inProc. EMNLP, 2020, pp. 6769–6781

    2020

  6. [6]

    Sentence-BERT: Sentence embeddings using Siamese BERT-networks,

    N. Reimers and I. Gurevych, “Sentence-BERT: Sentence embeddings using Siamese BERT-networks,” inProc. EMNLP-IJCNLP, 2019, pp. 3982–3992

    2019

  7. [7]

    Text embeddings by weakly-supervised contrastive pre- training,

    L. Wang, N. Yang, X. Huang, B. Jiao, L. Yang, D. Jiang, R. Majumder, and F. Wei, “Text embeddings by weakly-supervised contrastive pre- training,” arXiv:2212.03533, 2022

    Pith/arXiv arXiv 2022

  8. [8]

    Retrieval-augmented generation for knowledge-intensive NLP tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. K¨uttler, M. Lewis, W. Yih, T. Rockt ¨aschel, S. Riedel, and D. Kiela, “Retrieval-augmented generation for knowledge-intensive NLP tasks,” inAdv. Neural Inf. Process. Syst. (NeurIPS), 2020

    2020

  9. [9]

    IR-RAG@SIGIR25: The second edition of the workshop on information retrieval’s role in RAG systems,

    N. Arabzadeh, Z. Chen, F. Petroni, F. Siciliano, F. Silvestri, and G. Trappolini, “IR-RAG@SIGIR25: The second edition of the workshop on information retrieval’s role in RAG systems,” inProc. 48th Int. ACM SIGIR Conf. Research and Development in Information Retrieval, 2025

    2025

  10. [10]

    M3TR: Temporal retrieval enhanced multi-modal micro-video popularity prediction,

    J. Lu, W. Wang, M. Xiao, Y . Hua, T. Song, B. Peng, C. Hua, and H. Guan, “M3TR: Temporal retrieval enhanced multi-modal micro-video popularity prediction,” inProc. ACM Int. Conf. Multimedia (MM), 2026

    2026

  11. [11]

    Sparse, dense, and attentional representations for text retrieval,

    Y . Luan, J. Eisenstein, K. Toutanova, and M. Collins, “Sparse, dense, and attentional representations for text retrieval,”Trans. Assoc. Comput. Linguist., vol. 9, pp. 329–345, 2021

    2021

  12. [12]

    ColBERT: Efficient and effective passage search via contextualized late interaction over BERT,

    O. Khattab and M. Zaharia, “ColBERT: Efficient and effective passage search via contextualized late interaction over BERT,” inProc. 43rd Int. ACM SIGIR Conf., 2020, pp. 39–48

    2020

  13. [13]

    Reciprocal rank fusion outperforms Condorcet and individual rank learning methods,

    G. V . Cormack, C. L. A. Clarke, and S. B ¨uttcher, “Reciprocal rank fusion outperforms Condorcet and individual rank learning methods,” in Proc. 32nd Int. ACM SIGIR Conf., 2009, pp. 758–759

    2009

  14. [14]

    Item response theory based ensemble in machine learning,

    Z. Chen and H. Ahn, “Item response theory based ensemble in machine learning,”Int. J. Autom. Comput., vol. 17, no. 5, pp. 621–636, 2020

    2020

  15. [15]

    Query-Adaptive Hybrid Search,

    P. Posokhov, S. Skrylnikov, S. Masliukhin, A. Zavgorodniaia, O. Koro- teeva, and Y . Matveev, “Query-Adaptive Hybrid Search,”Mach. Learn. Knowl. Extr., vol. 8, no. 4, art. 91, 2026, doi:10.3390/make8040091

    work page doi:10.3390/make8040091 2026

  16. [16]

    GRASP: Plan-guided graph retrieval with adaptive fusion and reranking on semi- structured knowledge bases,

    Y . Tao, Y . Wang, X. Song, X. Luo, K. Liu, and J. Liu, “GRASP: Plan-guided graph retrieval with adaptive fusion and reranking on semi- structured knowledge bases,” arXiv:2605.30237, 2026

    Pith/arXiv arXiv 2026

  17. [17]

    When should active RAG retrieve? A budget-aware evaluation of utility, calibration, and cost,

    P. Qian, S. Wang, C. Peng, J. You, L. Liu, H. Yu, Y . Chen, and X. Jiang, “When should active RAG retrieve? A budget-aware evaluation of utility, calibration, and cost,” arXiv:2607.24010, 2026

    Pith/arXiv arXiv 2026

  18. [18]

    FinDER: Financial Dataset for Question Answering and Evaluating Retrieval-Augmented Generation,

    C. Choi, J. Kwon, J. Ha, H. Choi, C. Kim, Y . Lee, J.-Y . Sohn, and A. Lopez-Lira, “FinDER: Financial Dataset for Question Answering and Evaluating Retrieval-Augmented Generation,” inProc. 6th ACM Int. Conf. on AI in Finance (ICAIF), 2025, doi:10.1145/3768292.3770361

    arXiv 2025

  19. [19]

    Improving the completeness and com- parability of segment disclosures: A large language model approach,

    Y . Liu, Z. Cheng, and L. Lai, “Improving the completeness and com- parability of segment disclosures: A large language model approach,” arXiv:2605.23924, 2026

    Pith/arXiv arXiv 2026

  20. [20]

    From knowing to doing: A memory-controlled benchmark for LLM trading agents on stock markets,

    T. Zhu, W. Zhao, R. Sun, B. Luan, J. Lu, S. Wang, J. Li, D. Jiang, Y . He, and Z. Bai, “From knowing to doing: A memory-controlled benchmark for LLM trading agents on stock markets,” arXiv:2605.28359, 2026

    Pith/arXiv arXiv 2026

  21. [21]

    The efficiency frontier: A unified framework for cost-performance optimization in LLM context management,

    B. Shen, L. Jin, H. Cai, L. Hu, and Y . Xin, “The efficiency frontier: A unified framework for cost-performance optimization in LLM context management,” arXiv:2605.23071, 2026

    Pith/arXiv arXiv 2026

  22. [22]

    Billion-scale similarity search with GPUs,

    J. Johnson, M. Douze, and H. J ´egou, “Billion-scale similarity search with GPUs,”IEEE Trans. Big Data, vol. 7, no. 3, pp. 535–547, 2021

    2021

  23. [23]

    DriftGuard: Safety-aware multi-monitor detection and selective adaptation for evolving toxicity moderation,

    Y . Xin, H. Cai, B. Shen, L. Jin, and L. Hu, “DriftGuard: Safety-aware multi-monitor detection and selective adaptation for evolving toxicity moderation,” arXiv:2606.28725, 2026

    Pith/arXiv arXiv 2026

  24. [24]

    Random forests,

    L. Breiman, “Random forests,”Mach. Learn., vol. 45, no. 1, pp. 5–32, 2001

    2001

  25. [25]

    A comprehensive analysis of indicator effect on LLM performance,

    A. Ainiwaer, Q. Liu, and M. Lily, “A comprehensive analysis of indicator effect on LLM performance,” 2026, doi:10.13140/RG.2.2.29554.98248

    arXiv 2026

  26. [26]

    Scikit-learn: Machine learning in Python,

    F. Pedregosa et al., “Scikit-learn: Machine learning in Python,”J. Mach. Learn. Res., vol. 12, pp. 2825–2830, 2011

    2011

  27. [27]

    Local MDI+: Local feature importances for tree-based models,

    Z. Liang, Z. T. Rewolinski, A. Agarwal, T. M. Tang, and B. Yu, “Local MDI+: Local feature importances for tree-based models,”Trans. Mach. Learn. Res., 2026

    2026

  28. [28]

    Agentic AI as a cybersecurity attack surface: Threats, exploits, and defenses in runtime supply chains,

    X. Jiang, S. Yang, W. Yang, Y . Liu, and C. Ji, “Agentic AI as a cybersecurity attack surface: Threats, exploits, and defenses in runtime supply chains,” inProc. IEEE Conf. Artificial Intelligence (CAI), 2026, pp. 2142–2149

    2026

  29. [29]

    Revealing treatment non-adherence bias in clinical machine learning using large language models,

    Z. Liang, A. Suresh, and I. Y . Chen, “Revealing treatment non-adherence bias in clinical machine learning using large language models,” inProc. Conf. Health, Inference, and Learning (CHIL), 2025

    2025

  30. [30]

    AI for auto-research: Roadmap and user guide,

    L. Kong, X. Sun, W. Chow, L. Li, K. Q. Lin, X. B. Zhang, et al., “AI for auto-research: Roadmap and user guide,” arXiv:2605.18661, 2026

    Pith/arXiv arXiv 2026

  31. [31]

    DP-Prox: A robust and differentially private framework for federated instruction tuning of small LLMs on 8 GB edge devices,

    V . Venkatesh and Y . Liu, “DP-Prox: A robust and differentially private framework for federated instruction tuning of small LLMs on 8 GB edge devices,” inProc. IEEE 16th Annu. Computing and Communication Workshop and Conf. (CCWC), 2026, pp. 620–626

    2026

  32. [32]

    How LLMs work efficiently? A comprehensive study,

    Q. Liu, A. Ainiwaer, and Z. Lin, “How LLMs work efficiently? A comprehensive study,” 2026, doi:10.13140/RG.2.2.19111.00166

    arXiv 2026