Pith. sign in

REVIEW 4 major objections 5 minor 28 references

AI research idea systems can keep each researcher well-matched while stopping a whole community from being handed the same themes.

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 · grok-4.5

2026-07-31 18:12 UTC pith:EPXUMWC3

load-bearing objection Solid systems paper on a real portfolio problem; diversity gains look real, the “99.9% fit retained” half is mostly an internal LLM score. the 4 major comments →

arxiv 2607.28087 v1 pith:EPXUMWC3 submitted 2026-07-30 cs.AI

Diversifying Personalized Research Ideation against AI-Induced Homogenization

Rui Xu , Yunke Wang , Linwei Tao , Wenjie Xuan , Yong Luo This is my paper
classification cs.AI
keywords AI-assisted research ideationportfolio homogenizationpersonalized research directionsresearcher–direction alignmentcommunity-aware selectiondiversityLLM-as-judgeMMR
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.

When AI tools suggest research directions one researcher at a time, two failures show up. Coarse profiles push everyone toward the same mainstream ideas that look feasible but are only loosely tied to what a given person can actually do. And even when suggestions are individually reasonable, independent top picks can pile a whole field onto a few high-probability themes. DivAlign answers with a four-stage, training-free pipeline: build fine-grained profiles (lineage, owned artifacts, known gaps), generate a local pool of profile-conditioned directions, score each on Executability, Comprehensibility, and Growth Potential, then greedily pick one direction per researcher while penalizing near-duplicates already chosen for the community. On a 95-researcher AI benchmark across five subfields, that selection cuts portfolio redundancy versus both coarse single-shot ideation and independent top-choice, while keeping almost all of the measured researcher–direction fit.

Core claim

Portfolio-level homogenization is a real failure mode of AI-assisted research ideation, and it can be reduced without treating directions as interchangeable. By generating researcher-local candidate pools and selecting with an alignment score minus a max-similarity redundancy penalty, DivAlign lowers community nearest-neighbor and average pairwise embedding similarity relative to coarse single-shot and independent top-choice baselines, while retaining essentially the full mean alignment score of independent selection (about 99.9% at the default trade-off).

What carries the argument

DivAlign’s community-aware selection: from each researcher’s own scored candidate pool, greedily choose the pair that maximizes A(r,d) − λ·max cosine similarity to already surfaced directions (MMR-style max-sim redundancy), so every surfaced idea stays inside that researcher’s pool while near-duplicate exposure across the portfolio is penalized.

Load-bearing premise

The method treats three-part LLM judges of fit—from auto-extracted profiles—as a trustworthy stand-in for whether a real researcher can execute, critically own, and grow into a direction.

What would settle it

If independent experts who know the same researchers rate DivAlign’s assigned directions as no better matched (or worse) than independent top-choice, or if human portfolio comparisons show no drop in near-duplicate themes once embedding similarity is set aside, the alignment-preserving de-homogenization claim fails.

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

If this is right

  • Ideation systems should be judged on community portfolio redundancy, not only on single-idea quality or generation-side diversity.
  • Fine-grained profiles alone are not enough: multiple local candidates plus explicit cross-researcher selection are needed to cut repeated themes.
  • A modest redundancy weight can cut nearest-neighbor overlap while leaving measured fit almost unchanged versus picking each researcher’s best local score.
  • Paper-centric or multi-agent generators can plug into the candidate stage without changing the portfolio selection objective.

Where Pith is reading between the lines

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

  • Labs or funding programs that route many people through the same ideation model may silently correlate agendas unless they add a portfolio step like this.
  • Because human raters here judged coverage and distinctness—not personal fit—the practical next test is whether recipients accept and pursue the de-homogenized picks.
  • Low agreement between scorers on which local candidate ranks first, with similar portfolio metrics, suggests several near-equally fit alternatives often exist; selection mainly chooses among them.
  • The same alignment-plus-max-sim pattern could apply wherever personalized AI recommendations risk collective monoculture (grant themes, course projects, product roadmaps).

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 identifies portfolio-level homogenization in AI-assisted research ideation: independently generated directions can be individually plausible yet collectively redundant. It proposes DivAlign, a training-free four-stage pipeline that (1) extracts fine-grained researcher profiles (lineage, artifacts, gaps), (2) generates K profile-conditioned candidate directions, (3) scores each candidate with a three-component alignment rubric A=(E+C+G)/3 for Executability, Comprehensibility, and Growth Potential (Eq. 1), and (4) selects one direction per researcher via an MMR-style greedy rule that penalizes max cosine similarity to already chosen directions (Eqs. 2–3, Alg. 1). On a new benchmark of 95 AI researchers across five subfields, DivAlign (λ=0.2) reduces mean pairwise similarity HS from 0.331 to 0.294 and nearest-neighbor similarity NS from 0.704 to 0.608 versus coarse single-shot ideation, and reduces NS from 0.663 to 0.608 versus independent top-choice while reporting 99.9% retention of the mean Align. score (0.773 vs 0.774; Table 3). Ablations cover design progression, scoring signal, penalty form, λ, K, generator strength, community size, cross-scorer stability, and a blind human preference study on portfolio coverage/distinctness/quality (Fig. 5).

Significance. If the alignment-preserving claim holds, the work usefully reframes AI research ideation from generation-side novelty to recipient-side portfolio construction—a complementary problem to systems such as ResearchAgent, Nova, and ResearchTown. Strengths include a clear problem formulation, a concrete and reproducible pipeline (code/data promised), a multi-subfield 95-researcher benchmark with documented construction, and systematic ablations (Tables 4–9, Figs. 3–4) that isolate selection, penalty choice, and scaling with N. The pilot study (Tables 1–2) motivates the design without overclaiming collapse. The contribution is primarily methodological and empirical rather than theoretical; its lasting value depends on whether researcher–direction fit is measured in a way that is not largely internal to the same LLM pipeline used for generation.

major comments (4)
  1. [Table 3; Eq. 1; Stage 2–3; Human Evaluation / Fig. 5] The headline claim of “alignment-preserving” de-homogenization (Abstract; Table 3: Align. 0.774→0.773, “99.9% retained”) rests almost entirely on Stage-3 LLM scores A=(E+C+G)/3 over auto-extracted profiles (Eq. 1). Stage 2 already instructs the generator to satisfy executable / comprehensible / growth-enabling conditions using the same three notions, so Align. substantially re-scores the generator’s own brief. Human evaluation (Fig. 5; Human Evaluation section) explicitly does not assess researcher-specific fit—only portfolio Coverage, Distinctness, and Quality. Without an external fit check (e.g., self-ratings by profiled researchers, or blinded expert match of directions to held-out CVs), the 99.9% retention figure cannot support the central “alignment-preserving” claim even if embedding diversity gains are real. Please add such validation for a non-trivial subset, or substantially qua
  2. [Table 9; Additional Experiments / Cross-Scorer Validation] Cross-scorer validation (Table 9) shows Haiku vs Sonnet top-ranked candidate overlap of only 30.5% and Stage-4 assignment agreement of 35.8%, yet portfolio HS/NS remain similar. This is evidence of many near-tied local candidates and of judge instability at the researcher level—the level at which “fit preservation” is claimed. The paper should report uncertainty on Align. under scorer swap for the full portfolio (not only the 61 disagreed researchers), confidence intervals or bootstrap over researchers for the 0.001 Align. gap, and preferably a third independent judge family. As written, the 99.9% figure reads as an artifact of a smooth, same-family score landscape rather than a robust demonstration that real fit is preserved.
  3. [Table 3; Table 5; Related Works / Positioning] Baselines under-stress the portfolio claim. Main comparisons are Coarse-K1, Random, Independent (λ=0), and Diversity-Only (λ→∞) (Table 3). There is no strong generation-side diversity method (e.g., Nova-style iterative novelty search, explicit DPP/MMR at generation time, or multi-sample temperature/diverse-beam pools) nor a simple embedding-only de-duplication baseline that ignores E/C/G. Table 5’s TF-IDF cosine variant is useful but still uses DivAlign’s candidate pools and selection shell. To show that three-component alignment scoring plus community-aware selection is necessary—not just “more candidates + any diversity reranker”—please add at least one competitive generation-diversity baseline and one fit-agnostic portfolio diversifier on the same pools.
  4. [Metrics; Table 3; Pilot Study / Table 1] Diversity metrics HS/NS/VS are all SBERT (all-mpnet-base-v2) cosine geometry (Metrics; pilot near-duplicate threshold 0.8). Gains are modest in absolute terms (Independent→DivAlign: HS 0.303→0.294, NS 0.663→0.608). The paper should show that reductions are not embedding-specific (e.g., alternate encoder, lexical/n-gram diversity, or human-labeled near-duplicate rate on a sample) and should discuss practical significance: does NS 0.608 vs 0.663 correspond to fewer thematically repeated agendas a community would recognize? Linking a sample of high-NS pairs to qualitative themes (as in the pilot’s uncertainty-quantification example) would make the portfolio claim more convincing.
minor comments (5)
  1. [Eq. 1; Stage 3; Table 5] Growth Potential is described as an inverted-U / ZPD construct, but E, C, and G are averaged with equal weight (Eq. 1) with no sensitivity to unequal weights beyond the single-component ablations in Table 5. A brief weight sweep would help.
  2. [Figure 3; Stage 4] Figure 3’s λ sweep is helpful; please also report the fraction of researchers for whom DivAlign does not select the local top-A candidate, and the mean Align. drop conditional on those swaps—this directly explains the 0.001 aggregate gap.
  3. [Table 3; Experimental Setup] Table 3 reports E/C/G for Coarse-K1 after scoring against fine profiles “for fair evaluation,” which is reasonable but should be flagged more visibly in the table caption so readers do not treat Coarse-K1 Align. as same-protocol as the fine-pool methods.
  4. [Abstract; Introduction] Minor writing/spacing artifacts appear throughout (e.g., “nowcapableof”, “researchdirectionsconditionedon”, “retaining99.9%”). A full copy-edit pass is needed before camera-ready.
  5. [Data Usage and Ethics] Ethics note on using named researchers’ public profiles is appropriate; consider stating more explicitly that generated directions are not recommendations to those individuals and that release will include an opt-out or anonymization option if required by venue policy.

Circularity Check

2 steps flagged

Mild by-construction fit-retention under small-λ MMR; HS/NS diversity gains are independently measured, not circular.

specific steps
  1. self definitional [Stage 4 Eq. (3); Abstract; Table 3 Ours(λ=0) vs Ours(λ=0.2)]
    "At each greedy step, Stage 4 selects (i*, d*) = arg max_{i∈U, d∈Di} [A(ri, d)−λρ(d,S)] ... λ=0 recovers independent local top-choice selection. ... Compared with the independent top-choice variant, DivAlign reduces nearest-neighbor similarity from 0.663 to 0.608 while retaining 99.9% of the researcher-direction fit score."

    Researcher–direction fit is defined as A, and selection maximizes A minus a small redundancy penalty (default λ=0.2). Reporting that mean A under λ=0.2 is 99.9% of mean A under λ=0 is then largely forced by the objective and the existence of multiple high-A candidates in each Di, not an independent external demonstration that real fit is preserved. The diversity side (ρ via embeddings) is separate; only the fit-retention half is near-tautological for small λ.

  2. other [Stage 2 generation prompt; Stage 3 Eq. (1) and scoring prompt]
    "The generation prompt asks each candidate direction to satisfy three researcher-specific alignment conditions: executable... comprehensible... and growth-enabling... For researcher ri with profile pi and candidate direction d, the alignment score is A(ri, d)=1/3[E(ri,d)+C(ri,d)+G(ri,d)]."

    Stage 2 is instructed to emit directions that already meet Executability, Comprehensibility, and Growth Potential; Stage 3 re-scores the same three notions with the same model family. Absolute Align. therefore partly measures generator compliance with its own brief. This inflates and couples the fit metric used to claim “alignment-preserving,” though relative A comparisons within a fixed pool remain well-defined and HS/NS are not computed from E/C/G.

full rationale

DivAlign is an engineering pipeline paper, not a first-principles derivation. Community redundancy metrics (HS, NS, VS) are cosine/Vendi quantities on Sentence-BERT embeddings of surfaced titles/proposals and are not defined in terms of the alignment score A, so the reported drops (e.g., NS 0.663→0.608, HS 0.303→0.294 vs Independent; HS 0.331→0.294 vs Coarse-K1) are not circular. Human portfolio preferences (Coverage/Distinctness/Quality) are an external check. The only mild circularity is on the load-bearing “retains 99.9% fit” half of the headline: fit is operationalized as A=(E+C+G)/3 from Stage 3, Stage 2 already instructs the generator to satisfy the same three conditions, and Stage 4 explicitly maximizes A−λρ with small default λ=0.2—so near-full retention of mean A versus λ=0 is largely what the selector is built to do when multiple high-A local candidates exist (as cross-scorer disagreement also suggests). That is a weak self-definitional / by-construction loop on the alignment-preservation claim, not a collapse of the diversity results. No self-citation uniqueness chain, fitted-parameter-as-prediction, or renamed known theorem is load-bearing. Score 3 reflects one partial by-construction element around Align. retention, not central equivalence of the de-homogenization metrics to their inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

The load-bearing claim rests on operational definitions of fit and redundancy, hand-chosen trade-off settings, and the premise that public publication slices plus LLM extraction/scoring stand in for real researchers and real idea distinctness. No physical constants; the free parameters are pipeline knobs. Invented entities are methodological constructs (pipeline, rubric, benchmark), not ontological posits.

free parameters (4)
  • λ (redundancy weight) = 0.2 (default); sweep over {0, 0.2, 0.5, 1, 2, ∞}
    Controls A(r,d)−λρ trade-off; default λ=0.2 chosen as conservative balance from a sensitivity sweep, not derived.
  • K (candidates per researcher) = 5 (default); explored in {1,3,5,7,9}
    Size of local pool in Stage 2; default 5 from cost/diversity table, not theoretically fixed.
  • near-duplicate linking threshold (pilot) = 0.8
    Cosine ≥ 0.8 merges idea groups in the pilot Distinct metric, following prior work.
  • equal weights on E, C, G = 1/3 each
    A=(E+C+G)/3 hard-codes equal importance of the three rubric axes without learned or validated weights.
axioms (5)
  • domain assumption Sentence-embedding cosine similarity (all-mpnet-base-v2) is an adequate proxy for semantic redundancy among research directions.
    HS, NS, VS, and ρ(d,S) all depend on this; used throughout experiments and Stage 4.
  • domain assumption An LLM judge can score Executability, Comprehensibility, and Growth Potential from a short extracted profile in a way that tracks true researcher–direction fit.
    Stage 3 and all Align. columns rest on this; human study deliberately omits personalized fit.
  • domain assumption Up to 15 papers (2018–2022) plus homepage bio suffice to represent a researcher’s lineage, artifacts, and gaps for ideation.
    Benchmark construction and Stage 1 inputs; authors note profiles are partial.
  • ad hoc to paper Max pairwise similarity to the already chosen set is the right redundancy penalty for ‘repeated idea exposure.’
    Chosen over mean-sim because near-duplicates drive the failure mode; ablation supports NS/VS but is a design choice.
  • domain assumption Growth Potential is an inverted-U / Zone of Proximal Development function of skill overlap.
    Built into the Stage 3 rubric via Vygotsky-style guidance; shapes which candidates score high on G.
invented entities (3)
  • DivAlign four-stage pipeline no independent evidence
    purpose: Alignment-preserving community portfolio selection over researcher-local idea pools.
    Core proposed method; evaluated only inside this paper’s benchmark and judges.
  • E/C/G researcher–direction fit rubric no independent evidence
    purpose: Decompose alignment beyond topical match into executability, comprehensibility, growth potential.
    New scoring object used as both generation constraint and selection utility; no external validated instrument.
  • 95-researcher multi-subfield ideation benchmark no independent evidence
    purpose: Provide a community-scale testbed for homogenization and fit metrics.
    Constructed artifact from public metadata; valuable if released, but not an independent natural experiment.

pith-pipeline@v1.2.0-daily-grok45 · 21453 in / 4025 out tokens · 90914 ms · 2026-07-31T18:12:00.588319+00:00 · methodology

0 comments
read the original abstract

AI-assisted research ideation has emerged as a promising paradigm for accelerating scientific discovery, with systems now capable of generating research directions conditioned on papers, topics, or lightweight researcher contexts. Yet current systems largely optimize individual suggestions in isolation. This leaves two blind spots. First, coarse researcher representations may elicit mainstream directions that appear broadly feasible, but lack sufficient researcher-specific grounding. Second, independent recommendations can concentrate a community's portfolio around recurring high-probability themes. To address these blind spots, we propose DivAlign, a four-stage pipeline for alignment-preserving de-homogenization. DivAlign extracts fine-grained researcher profiles, generates profile-conditioned candidate directions, scores them along three alignment dimensions (Executability, Comprehensibility, and Growth Potential), and surfaces researcher-local directions while reducing redundancy across the community portfolio. On a benchmark we construct from 95 AI researchers across five subfields, DivAlign reduces community-level redundancy while preserving researcher-direction fit. Compared with coarse single-shot ideation, it lowers average pairwise similarity from 0.331 to 0.294 and nearest-neighbor similarity from 0.704 to 0.608. Compared with the independent top-choice variant, DivAlign reduces nearest-neighbor similarity from 0.663 to 0.608 while retaining 99.9% of the researcher-direction fit score. Code and data are available at https://github.com/Ruixxxx/DivAlign.

Figures

Figures reproduced from arXiv: 2607.28087 by Linwei Tao, Rui Xu, Wenjie Xuan, Yong Luo, Yunke Wang.

Figure 1
Figure 1. Figure 1: (a) Simple profile-conditioned ideation can still [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of DivAlign. The colored researcher cards on the left denote a set of researchers [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sensitivity to the redundancy weight λ. Selection Generator HS↓ NS↓ VS↑ Align.↑ Independent Haiku 0.303 0.663 0.262 0.779 Independent Sonnet 0.319 0.689 0.244 0.784 DivAlign Haiku 0.294 0.608 0.289 0.776 DivAlign Sonnet 0.296 0.620 0.282 0.785 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Human evaluation results. Human Evaluation We further conduct a blind pairwise human evaluation. Since researcher-specific alignment is difficult for external eval￾uators to assess, we focus on portfolio-level diversity and perceived quality. Evaluators are asked to compare cluster￾specific portfolio subsets (5 directions each) on Coverage (breadth of covered research directions), Distinctness (less intern… 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

28 extracted references · 3 linked inside Pith

  1. [1]

    Towards End-to-End Automation of

    Lu, Chris and Lu, Cong and Lange, Robert Tjarko and Yamada, Yutaro and Hu, Shengran and Foerster, Jakob and Ha, David and Clune, Jeff , journal =. Towards End-to-End Automation of. 2026 , doi =

    2026

  2. [2]

    Yamada, Yutaro and Lange, Robert Tjarko and Lu, Cong and Hu, Shengran and Lu, Chris and Foerster, Jakob and Clune, Jeff and Ha, David , journal =. The

  3. [3]

    2025 , doi =

    Baek, Jinheon and Jauhar, Sujay Kumar and Cucerzan, Silviu and Hwang, Sung Ju , booktitle =. 2025 , doi =

    2025

  4. [4]

    Lyu, Yougang and Zhang, Xi and Yi, Xinhao and Zhao, Yuyue and Guo, Shuyu and Hu, Wenxiang and Piotrowski, Jan and Kaliski, Jakub and Urbani, Jacopo and Meng, Zaiqiao and Zhou, Lun and Yan, Xiaohui , journal =

  5. [5]

    and Hauser, Oliver P

    Doshi, Anil R. and Hauser, Oliver P. , journal =. Generative. 2024 , doi =

    2024

  6. [6]

    arXiv preprint arXiv:2604.09502 , year =

    Strategic Algorithmic Monoculture: Experimental Evidence from Coordination Games , author =. arXiv preprint arXiv:2604.09502 , year =

    Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2512.15011 , year =

    Epistemic Diversity Across Language Models Mitigates Knowledge Collapse , author =. arXiv preprint arXiv:2512.15011 , year =

    Pith/arXiv arXiv

  8. [8]

    Transactions on Machine Learning Research , year =

    The Vendi Score: A Diversity Evaluation Metric for Machine Learning , author =. Transactions on Machine Learning Research , year =

  9. [9]

    ICLR , year =

    Does Writing with Language Models Reduce Content Diversity? , author =. ICLR , year =

  10. [10]

    EACL , pages =

    Evaluating the Evaluation of Diversity in Natural Language Generation , author =. EACL , pages =. 2021 , doi =

    2021

  11. [11]

    Mind in Society: The Development of Higher Psychological Processes , author =

  12. [12]

    Sentence-

    Reimers, Nils and Gurevych, Iryna , booktitle =. Sentence-. 2019 , doi =

    2019

  13. [13]

    The Use of

    Carbonell, Jaime and Goldstein, Jade , booktitle =. The Use of

  14. [14]

    Many Heads Are Better Than One: Improved Scientific Idea Generation by a

    Su, Haoyang and Chen, Renqi and Tang, Shixiang and Yin, Zhenfei and Zheng, Xinzhe and Li, Jinzhe and Qi, Biqing and Wu, Qi and Li, Hui and Ouyang, Wanli and Torr, Philip and Zhou, Bowen and Dong, Nanqing , booktitle =. Many Heads Are Better Than One: Improved Scientific Idea Generation by a. 2025 , doi =

    2025

  15. [15]

    arXiv preprint arXiv:2506.18348 , year =

    Dynamic Knowledge Exchange and Dual-Diversity Review: Concisely Unleashing the Potential of a Multi-Agent Research Team , author =. arXiv preprint arXiv:2506.18348 , year =

    Pith/arXiv arXiv

  16. [16]

    Si, Chenglei and Yang, Diyi and Hashimoto, Tatsunori , booktitle =. Can

  17. [17]

    Wang, Chengcheng and Xie, Qinhua and He, Wei and Guo, Jianyuan and Wang, Shiqi and Xu, Chang , journal =. Sibyl-

  18. [18]

    Yu, Haofei and Hong, Zhaochen and Cheng, Zirui and Zhu, Kunlun and Xuan, Keyang and Yao, Jinwei and Feng, Tao and You, Jiaxuan , booktitle =

  19. [19]

    Diversity Collapse in Multi-Agent

    Chen, Nuo and Tong, Yicheng and Yang, Yuzhe and He, Yufei and Zhang, Xueyi and Zou, Qingyun and Wang, Qian and He, Bingsheng , booktitle =. Diversity Collapse in Multi-Agent. 2026 , doi =

    2026

  20. [20]

    The Ideation-Execution Gap: Execution Outcomes of

    Si, Chenglei and Hashimoto, Tatsunori and Yang, Diyi , booktitle =. The Ideation-Execution Gap: Execution Outcomes of

  21. [21]

    2025 , doi =

    Hu, Xiang and Fu, Hongyu and Wang, Jinge and Wang, Yifeng and Li, Zhikun and Xu, Renjun and Lu, Yu and Jin, Yaochu and Pan, Lili and Lan, Zhenzhong , booktitle =. 2025 , doi =

    2025

  22. [22]

    2024 , doi =

    Wang, Qingyun and Downey, Doug and Ji, Heng and Hope, Tom , booktitle =. 2024 , doi =

    2024

  23. [23]

    Tang, Jiabin and Xia, Lianghao and Li, Zhonghang and Huang, Chao , booktitle =

  24. [24]

    arXiv preprint arXiv:2510.04226 , year =

    Epistemic Diversity and Knowledge Collapse in Large Language Models , author =. arXiv preprint arXiv:2510.04226 , year =

    arXiv

  25. [25]

    Task-Dependent Evaluation of

    Jain, Shomik and Lanchantin, Jack and Nickel, Maximilian and Ross, Candace and Ullrich, Karen and Wilson, Ashia and Watson-Daniels, Jamelle , journal =. Task-Dependent Evaluation of

  26. [26]

    Examining and Addressing Barriers to Diversity in

    Deng, Yuting and Brucks, Melanie and Toubia, Olivier , journal =. Examining and Addressing Barriers to Diversity in

  27. [27]

    Scientific Reasoning Does Not Reliably Translate into Scientific Forecasting in Frontier

    Wu, Sean and Lu, Pan and Chen, Yupeng and Bragg, Jonathan and Yamada, Yutaro and Clark, Peter and Clifton, David and Torr, Philip and Zou, James and Yu, Junchi , journal =. Scientific Reasoning Does Not Reliably Translate into Scientific Forecasting in Frontier

  28. [28]

    Deep Research Arena: The First Exam of

    Wan, Haiyuan and Yang, Chen and Yu, Junchi and Tu, Meiqi and Lu, Jiaxuan and Yu, Di and Cao, Jianbao and Gao, Ben and Xie, Jiaqing and Wang, Aoran and Zhang, Wenlong and Torr, Philip and Zhou, Dongzhan , booktitle =. Deep Research Arena: The First Exam of