Pith. sign in

REVIEW 3 major objections 5 minor 34 references

The paper argues that adding a pre-retrieval query profile and trace-conditioned hints to an adaptive knowledge-graph retriever consistently hurts final ranking quality under open-weight controllers, even though it makes the retrieval proce

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

2026-08-02 08:31 UTC pith:HEYLBMPP

load-bearing objection Cleanly designed negative result on profile/trace conditioning for adaptive KG retrieval, but an unlisted static control instruction in the prompt could confound the main finding. the 3 major comments →

arxiv 2607.24800 v1 pith:HEYLBMPP submitted 2026-07-06 cs.IR cs.AIcs.CL

When Thinking Before Retrieval Hurts: TraceBound Diagnostics for Adaptive Knowledge-Graph Retrieval

Partha Sarathi Purkayastha (ETH Z\"urich) This is my paper
classification cs.IR cs.AIcs.CL
keywords adaptive retrievalknowledge-graph question answeringtrace diagnosticsretrieval agentsquery profilingaction selectionfailure analysisretrieval-augmented generation
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.

This paper sets out to test a widespread intuition: that giving an adaptive retrieval agent explicit 'thinking' before it searches — a compact query profile and short hints derived from its own trace — should make retrieval better on text-rich knowledge graphs. The agent runs an action loop with two tools, broad search and neighborhood exploration, and is evaluated by standard ranking metrics. Across validation and held-out examples and in paired same-query comparisons, adding the profile and trace hints consistently lowers Hit@5, Recall@20, and MRR relative to the unconditioned controller, while providing richer diagnostic logs. The trace data show that the degradation is localized to repeated calls, zero-result calls, and exploration budget spent without rank improvement, and that a tighter interaction budget shortens trajectories without repairing the policy. The paper's conclusion is that 'thinking before retrieval' should be treated as a control problem over action selection — deciding when to search, traverse, relax, rerank, or stop — not as a prompt-format change.

Core claim

The paper's central claim is that, with graph data, tools, labels, and metrics fixed, adding a query-profile block and trace-conditioned hints to an adaptive retrieval controller consistently reduces final retrieval quality while improving inspectability. Validation, held-out, and paired same-query comparisons all show negative deltas in Hit@5, Recall@20, and MRR. The loss is largest on an academic-paper graph with dense relation neighborhoods. Trace analysis attributes the loss to repeated calls, zero-result calls, and exploration budget spent without rank gain; a constrained-budget probe shortens trajectories but does not restore ranking. The authors conclude that profile and trace informa

What carries the argument

The central object is TraceBound, a diagnostic protocol that wraps an adaptive retriever's action loop with three components: a query-profile block built before the first tool call (predicted answer type, facets, relation hints, risks); trace-conditioned hints issued after symptoms like repeated or zero-result calls, with a small vocabulary (re-anchor, relax, tighten, stop); and trajectory instrumentation that logs counters for global calls, neighborhood calls, zero-result calls, repeated calls, steps, and budget. The protocol fixes graph data, tools, labels, and evaluator so that any change in ranking metrics is attributable to the controller. It also separates ranking evaluation (computed

Load-bearing premise

The diagnosis rests on the assumption that the trace counters (especially repeated calls and zero-result calls) faithfully and causally capture the policy failures behind the rank loss; the constrained-budget probe lacks a paired unconditioned baseline, so the evidence cannot fully rule out that the conditioned agent simply thinks too long rather than chooses the wrong actions.

What would settle it

A direct check: run the unconditioned controller under the same reduced step/call budget as the constrained TraceBound probe, on the same held-out queries. If the unconditioned agent under that tight budget matches TraceBound's low MRR, then the loss is explained by budget exhaustion rather than wrong action choices; if it remains clearly better, the action-selection explanation is supported. A second check: train a small router that maps trace counters to the next action and compare it to both conditions; if it beats the unconditioned baseline, the negative result is specific to static, promp

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

If this is right

  • Adding profile and trace context to an adaptive graph-retrieval controller can degrade ranking even when it improves process transparency, so such context should not be treated as globally helpful.
  • Trace counters — repeated calls, zero-result calls, and exploration counts — predict larger rank loss and can expose failures that final metrics hide.
  • A tighter interaction budget shortens trajectories but does not recover accuracy, indicating the failure is about action selection rather than excess length.
  • Future adaptive retrieval should replace static prompt-level conditioning with a learned policy over trace state variables such as candidate coverage, repetition, and budget.
  • For personalization, injecting user-like context before retrieval should be gated by a conditional policy rather than unconditionally concatenated.

Where Pith is reading between the lines

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

  • If the result generalizes, personalized retrieval systems should avoid concatenating user-state text into every retrieval decision; the trace-state formulation suggests a learnable gating mechanism as a more robust route.
  • The conditional repeated-call result implies a testable hypothesis: explicitly penalizing or aborting duplicate actions during training may remove the largest chunk of the measured degradation, since repetition is where the rank loss concentrates.
  • The protocol's fixed invariants make it portable to other two-tool retrieval agents; whether the negative effect persists with stronger or differently calibrated base models remains an open empirical question, since all evidence here comes from locally served open-weight controllers.

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

3 major / 5 minor

Summary. The paper studies adaptive knowledge-graph retrieval with an ARK-style controller on STaRK, asking whether adding a pre-retrieval query profile and trace-conditioned hints (the 'TraceBound' intervention) improves retrieval quality. Across validation ablations (Qwen3-14B), held-out subsets (Qwen3-30B-A3B), and shared-ID paired comparisons on PRIME, MAG, and Amazon, the paper reports that the unconditioned controller consistently outperforms TraceBound on Hit@1, Hit@5, Recall@20, and MRR. Trace-counter analysis associates repeated calls and zero-result calls with larger rank drops, and a constrained-budget probe shows that shortening trajectories does not recover accuracy. The authors conclude that profile/trace conditioning improves inspectability but degrades retrieval, and that adapter design should be framed as a control problem over action selection.

Significance. If the negative result is valid, it is a useful cautionary contribution for the adaptive-retrieval and personalization communities: it provides a controlled, reproducible intervention that isolates controller-prompt changes from changes to data, tools, labels, and metrics, and it pairs ranking metrics with trajectory-level diagnostics. The shared-ID paired comparisons and win/tie/loss counts are good practice. However, the headline claim rests on the internal validity of the intervention, and the current manuscript leaves a plausible confound unaddressed; the lack of uncertainty quantification and the incompletely controlled budget probe also limit the strength of the conclusions. The paper is well positioned for a workshop audience, but the central claim requires an experimental fix.

major comments (3)
  1. [§3.2, Table 8, Appendix E] The intervention is defined as adding only a profile block and a trace hint (Table 8), but the TraceBound prompt skeleton in Appendix E contains a [CONTROL INSTRUCTION] block: 'Choose exactly one: Search, Explore, Answer, or Backtrack. Do not repeat an earlier call unless it is necessary. Prefer answering when additional calls are unlikely to improve rank.' This static instruction biases the controller toward early stopping and avoiding repeats. If the unconditioned baseline does not include this same instruction, the comparison changes two variables at once, and the observed MRR drops could stem from the control instruction alone. The paper neither states that the baseline receives the identical instruction nor releases the exact prompts. This confound threatens the existence of the central negative result, not merely its mechanism. Please either confirm that the baseline includes the s
  2. [Tables 3, 4, and Figure 3] The headline claim of a consistent negative effect is presented without any uncertainty quantification. On PRIME and Amazon the paired deltas are small (e.g., MRR deltas of -0.0653 and -0.0426 in Table 4), and the win/tie/loss counts show many ties (192 and 190 of 291/297 queries). The paper reports no confidence intervals, standard errors, or significance tests. Given that the contribution is an empirical negative result, the authors should report bootstrap CIs or permutation-test p-values for the paired deltas, or at least per-graph paired significance tests, to establish that the degradation is not within sampling noise. The validation ablation in Table 2 uses only 100 examples per graph, which further amplifies the need for interval estimates.
  3. [§6.2, Table 11] The constrained-budget probe is used to argue that the failure is not 'the agent thinks too long' but rather 'the agent chooses wrong actions.' However, Table 11 provides no paired unconditioned baseline under the same reduced budget and call caps. The comparison is against unconditioned runs with larger budgets (Tables 2 and 3), so it cannot separate the effect of the tighter budget from the effect of TraceBound's action policy. A proper probe should run the unconditioned controller with the same step budget and call caps on the same query IDs. As reported, the evidence does not support the claim that budget reduction fails to repair the policy, because any controller could fail in the same way under such tight constraints. This is a load-bearing point for RQ3, not just a presentation issue.
minor comments (5)
  1. [Appendix E] The statement 'The exact prompts are implementation artifacts' is not sufficient for reproducibility. Please release the full prompt templates for all four controller variants (unconditioned, profile-only, trace-only, TraceBound) in the appendix or supplementary material, so readers can verify the control instruction is held constant.
  2. [Abstract and §1] The phrase 'the common failure mode in that "thinking before retrieval"' is ungrammatical and obscures the intended claim. Please rephrase, e.g., 'diagnoses the common failure mode in which...'.
  3. [§4] The paper notes that 'long-running agent jobs can complete slightly different numbers of examples per condition' and then reports aggregate results on different n. This raises the question of whether incompletions are systematic (e.g., timeouts correlated with condition). Please report how many examples failed per condition and whether shared-ID paired analysis excludes them equally.
  4. [§5.1, Table 2] The 'Steps' column is labeled as maximum interaction budget in the text, but the table heading alone is ambiguous; consider renaming to 'Max steps' or clarifying in the caption that the reported step values are the budget limits, not averages.
  5. [§6.1] The claim that TraceBound 'improves inspectability' is never operationalized. Consider reporting a concrete measure, such as the fraction of trajectories with a logged repeat/zero event, or at least stating explicitly that inspectability is a qualitative benefit.

Circularity Check

0 steps flagged

No significant circularity: the central result is an external benchmark comparison, and trace diagnostics are post-hoc measurements rather than fitted predictions.

full rationale

The paper makes no derivation that reduces to its inputs. Its central claim—'the added conditioning improves inspectability but consistently reduces retrieval quality'—is an empirical comparison on STaRK with fixed graph data, tools, gold labels, and metrics (Table 8; §3.5). The unconditioned baseline and TraceBound are both run on the same external benchmark, and the reported deltas (Tables 2–4) are computed from final ranked lists. The trace counters in §6.1 (global, neighborhood, zero-result, repeated calls) are logged from trajectories, not fitted to the rank deltas; Figure 4 is a conditional association, not a predictive model with fitted parameters. The failure taxonomy (Table 6) is an interpretive summary of the same logs, but it is not used to generate the retrieval results, so it is not circular. There is no load-bearing self-citation: the cited ARK, STaRK, and Qwen3 works are by other authors and provide external architecture/benchmark/model context. The only notable caveat is internal validity, not circularity: Appendix E's [CONTROL INSTRUCTION] block ('Prefer answering when additional calls are unlikely to improve rank') is not listed in Table 8 as an added component, so if the unconditioned baseline did not receive the same instruction the comparison would be confounded. That is a correctness/experimental-design concern, not a case where the result is equivalent to its input by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical or mathematical entities; it introduces a protocol (TraceBound) composed of existing components (prompt blocks, trace logging). The main assumptions are representativeness of the ARK interface, validity of STaRK and its metrics, and fidelity of the trace counters as causal indicators.

axioms (5)
  • domain assumption The ARK-style two-tool interface (global lexical search + one-hop neighborhood exploration) is a representative instantiation of adaptive KG retrieval.
    The central claim is tested only on this interface (§3.1). If this interface is not representative, the negative result may not generalize to other adaptive retrieval systems.
  • domain assumption The STaRK benchmark's gold answers and graph structures are correct and sufficiently cover the target domains.
    All metrics are computed against STaRK gold labels (§4). Erroneous or noisy labels would bias the measured deltas.
  • standard math Hit@1, Hit@5, Recall@20, and MRR are valid measures of retrieval quality for the claims made.
    The paper defines these metrics in Appendix F and assumes they capture the construct of 'retrieval quality' used in the strongest claim.
  • domain assumption The trace counters (repeated calls, zero-result calls, budget) faithfully reflect the failure modes they are used to explain.
    The mechanism analysis in §6.1 correlates counters with rank deltas. If the counters are noisy or mis-specified, the diagnostic explanation could be spurious.
  • domain assumption Open-weight Qwen3 controllers are sufficiently representative of LLM-based retrieval agents to support the general conclusion about 'thinking before retrieval'.
    Only Qwen3-14B and Qwen3-30B-A3B are tested. The conclusion that the problem is a control problem rather than a prompt-format change is broader than the tested model family.

pith-pipeline@v1.3.0-alltime-deepseek · 246 in / 9580 out tokens · 131121 ms · 2026-08-02T08:31:09.899851+00:00 · methodology

0 comments
read the original abstract

Adaptive retrieval promises to make knowledge-graph question answering more robust by letting a controller search, inspect neighborhoods, revise actions, and stop when evidence is sufficient. We study this premise by introducing TraceBound, a lightweight profile- and trace-conditioned diagnostic protocol for an ARK-style retriever on text-rich knowledge graphs. TraceBound exposes a compact query profile before retrieval, issues short trace hints after observable failure symptoms, and logs trajectory counters, while keeping graph data, tools, gold labels, and ranking metrics fixed. Across STaRK validation and held-out subsets, the added conditioning improves inspectability but consistently reduces retrieval quality under open-weight controllers. Paired trajectory analysis localizes the degradation to repeated calls, zero-result calls, and misallocated exploration budget, while stricter interaction budgets shorten trajectories without repairing the policy. The result diagnoses the common failure mode in that "thinking before retrieval'' must be evaluated as a control problem over action selection, not as a prompt-format change.

Figures

Figures reproduced from arXiv: 2607.24800 by Partha Sarathi Purkayastha (ETH Z\"urich).

Figure 1
Figure 1. Figure 1: TraceBound workflow. The input query and graph are passed to an ARK-style controller. TraceBound adds a query-profile block before retrieval and a monitor that summarizes zero-result calls, repeated calls, budget use, and step count during retrieval. The graph tools and final ranking evaluator are unchanged. The ranked answers produce Hit@1, Hit@5, Recall@20, and MRR, while the monitor produces trace diagn… view at source ↗
Figure 2
Figure 2. Figure 2: Validation MRR by controller variant and step bud￾get. The dashed horizontal reference line marks the best uncon￾ditioned baseline. All profile/trace variants fall below that line, so the intervention improves neither the 12-step nor the 16-step validation setting. PRIME MAG Amazon 0 −0.05 −0.1 −0.15 −0.2 −0.25 TraceBound - Unconditioned Hit@5 Recall@20 MRR [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Held-out metric deltas. Each grouped bar is Trace￾Bound minus unconditioned on the held-out subset. Bars below zero mean the profile/trace intervention is worse. The three bars per graph correspond to Hit@5, Recall@20, and MRR; all are negative across PRIME, MAG, and Amazon. but wrong neighborhoods. 5.3. Paired comparison: the result is not an artifact of completed-count differences [PITH_FULL_IMAGE:figur… view at source ↗
Figure 4
Figure 4. Figure 4: Conditional rank loss by repeated-call bucket. Each bar is the mean paired MRR delta, TraceBound minus uncondi￾tioned. Here, “repeat” means at least one repeated or near-duplicate call occurred. Repetition predicts a larger loss on PRIME and MAG, and does not help on Amazon either. weak. This separates two hypotheses. If the main failure were only excessive length, tighter budgets should recover most of th… 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

34 extracted references · 4 linked inside Pith

  1. [1]

    Self- RAG : Learning to retrieve, generate, and critique through self-reflection

    Asai, A., Wu, Z., Wang, Y., Sil, A., and Hajishirzi, H. Self- RAG : Learning to retrieve, generate, and critique through self-reflection. In International Conference on Learning Representations, 2024

    2024

  2. [2]

    Autofocus retrieval: An effective pipeline for multi-hop question answering with semi-structured knowledge

    Boer, D., Roth, S., and Kramer, S. Autofocus retrieval: An effective pipeline for multi-hop question answering with semi-structured knowledge. Transactions on Machine Learning Research, 2026. arXiv:2505.09246

    Pith/arXiv arXiv 2026

  3. [3]

    and Shi, J

    Clemedtson, A. and Shi, J. GraphRAFT : Retrieval augmented fine-tuning for knowledge graphs on graph databases. arXiv preprint arXiv:2504.05478, 2025

    arXiv 2025

  4. [4]

    Dense passage retrieval for open-domain question answering

    Karpukhin, V., Oguz, B., Min, S., Lewis, P., Wu, L., Edunov, S., Chen, D., and Yih, W.-t. Dense passage retrieval for open-domain question answering. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2020

    2020

  5. [5]

    Retrieval-augmented generation for knowledge-intensive NLP tasks

    Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Kuttler, H., Lewis, M., Yih, W.-t., Rockt \"a schel, T., Riedel, S., and Kiela, D. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems, 2020

    2020

  6. [6]

    AgentBench : Evaluating LLM s as agents

    Liu, X., Yu, H., Zhang, H., Xu, Y., Lei, X., Lai, H., Gu, Y., Ding, H., Men, K., Yang, K., et al. AgentBench : Evaluating LLM s as agents. In International Conference on Learning Representations, 2024

    2024

  7. [7]

    Self-refine: Iterative refinement with self-feedback

    Madaan, A., Tandon, N., Gupta, P., Hallinan, S., Gao, L., Wiegreffe, S., Alon, U., Dziri, N., Prabhumoye, S., Yang, Y., et al. Self-refine: Iterative refinement with self-feedback. In Advances in Neural Information Processing Systems, 2023

    2023

  8. [9]

    Toolformer: Language models can teach themselves to use tools

    Schick, T., Dwivedi-Yu, J., Dess \`i , R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., and Scialom, T. Toolformer: Language models can teach themselves to use tools. In Advances in Neural Information Processing Systems, 2023

    2023

  9. [10]

    Reflexion: Language agents with verbal reinforcement learning

    Shinn, N., Cassano, F., Gopinath, A., Narasimhan, K., and Yao, S. Reflexion: Language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems, 2023

    2023

  10. [12]

    Personabench: Evaluating AI models on understanding personal information through accessing synthetic private user data

    Tan, J., Yang, L., Liu, Z., Liu, Z., Murthy, R., Awalgaonkar, T., Zhang, J., Yao, W., Zhu, M., Kokane, S., et al. Personabench: Evaluating AI models on understanding personal information through accessing synthetic private user data. In Findings of the Association for Computational Linguistics, 2025

    2025

  11. [13]

    Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions

    Trivedi, H., Balasubramanian, N., Khot, T., and Sabharwal, A. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. In Proceedings of the Association for Computational Linguistics, 2023

    2023

  12. [14]

    N., Subbian, K., Zou, J., and Leskovec, J

    Wu, S., Zhao, S., Yasunaga, M., Huang, K., Cao, K., Huang, Q., Ioannidis, V. N., Subbian, K., Zou, J., and Leskovec, J. STaRK : Benchmarking LLM retrieval on textual and relational knowledge bases. In Advances in Neural Information Processing Systems, Datasets and Benchmarks Track, 2024

    2024

  13. [15]

    J., Cheng, Z., Shin, D., Lei, F., et al

    Xie, T., Zhang, D., Chen, J., Li, X., Zhao, S., Cao, R., Hua, T. J., Cheng, Z., Shin, D., Lei, F., et al. OSWorld : Benchmarking multimodal agents for open-ended tasks in real computer environments. In Advances in Neural Information Processing Systems, 2024

    2024

  14. [17]

    ReAct : Synergizing reasoning and acting in language models

    Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., and Cao, Y. ReAct : Synergizing reasoning and acting in language models. In International Conference on Learning Representations, 2023

    2023

  15. [18]

    Personalize before retrieve: LLM -based personalized query expansion for user-centric retrieval

    Zhang, Y., Jia, P., Xu, D., Wen, Y., Li, X., Wang, Y., Zhang, W., Li, X., Gan, W., Guo, H., Liu, Y., and Zhao, X. Personalize before retrieve: LLM -based personalized query expansion for user-centric retrieval. arXiv preprint arXiv:2510.08935, 2025

    arXiv 2025

  16. [19]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and Kuttler, Heinrich and Lewis, Mike and Yih, Wen-tau and Rockt. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems , year =

  17. [20]

    Proceedings of the Conference on Empirical Methods in Natural Language Processing , year =

    Dense Passage Retrieval for Open-Domain Question Answering , author =. Proceedings of the Conference on Empirical Methods in Natural Language Processing , year =

  18. [21]

    Proceedings of the Association for Computational Linguistics , year =

    Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions , author =. Proceedings of the Association for Computational Linguistics , year =

  19. [22]

    Asai, Akari and Wu, Zeqiu and Wang, Yizhong and Sil, Avirup and Hajishirzi, Hannaneh , booktitle =. Self-

  20. [23]

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle =

  21. [24]

    Advances in Neural Information Processing Systems , year =

    Reflexion: Language Agents with Verbal Reinforcement Learning , author =. Advances in Neural Information Processing Systems , year =

  22. [25]

    Advances in Neural Information Processing Systems , year =

    Self-Refine: Iterative Refinement with Self-Feedback , author =. Advances in Neural Information Processing Systems , year =

  23. [26]

    Advances in Neural Information Processing Systems , year =

    Toolformer: Language Models Can Teach Themselves to Use Tools , author =. Advances in Neural Information Processing Systems , year =

  24. [27]

    arXiv preprint arXiv:2307.07697 , year =

    Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph , author =. arXiv preprint arXiv:2307.07697 , year =

    Pith/arXiv arXiv

  25. [28]

    and Subbian, Karthik and Zou, James and Leskovec, Jure , booktitle =

    Wu, Shirley and Zhao, Shichang and Yasunaga, Michihiro and Huang, Kaidi and Cao, Kaidi and Huang, Qian and Ioannidis, Vassilis N. and Subbian, Karthik and Zou, James and Leskovec, Jure , booktitle =

  26. [29]

    Proceedings of the ACM SIGOPS Symposium on Operating Systems Principles , year=

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author=. Proceedings of the ACM SIGOPS Symposium on Operating Systems Principles , year=

  27. [30]

    Transactions on Machine Learning Research , year=

    Autofocus Retrieval: An Effective Pipeline for Multi-Hop Question Answering With Semi-Structured Knowledge , author=. Transactions on Machine Learning Research , year=

  28. [31]

    Clemedtson, Anton and Shi, Jiaxin , journal =

  29. [32]

    arXiv preprint arXiv:2601.13969 , year =

    Autonomous Knowledge Graph Exploration with Adaptive Breadth-Depth Retrieval , author =. arXiv preprint arXiv:2601.13969 , year =

    Pith/arXiv arXiv

  30. [33]

    Personalize Before Retrieve:

    Zhang, Yixuan and Jia, Pengyue and Xu, Dongjing and Wen, Yiming and Li, Xin and Wang, Yuyang and Zhang, Wenjie and Li, Xiaoyan and Gan, Wensheng and Guo, Huifeng and Liu, Yong and Zhao, Xiangyu , journal =. Personalize Before Retrieve:

  31. [34]

    PersonaBench: Evaluating

    Tan, Jiachen and Yang, Liu and Liu, Zihan and Liu, Zhiwei and Murthy, Raghavendra and Awalgaonkar, Tulika and Zhang, Jiawei and Yao, Weiran and Zhu, Ming and Kokane, Sairam and others , booktitle =. PersonaBench: Evaluating

  32. [35]

    Liu, Xiao and Yu, Hao and Zhang, Hanchen and Xu, Yifan and Lei, Xuanyu and Lai, Hanyu and Gu, Yuxian and Ding, Hangliang and Men, Kaiwen and Yang, Kejuan and others , booktitle =

  33. [36]

    Xie, Tianbao and Zhang, Danyang and Chen, Jixuan and Li, Xiaochuan and Zhao, Siyuan and Cao, Ruisheng and Hua, T. J. and Cheng, Zesen and Shin, Dongyeop and Lei, Fangyu and others , booktitle =

  34. [37]

    arXiv preprint arXiv:2505.09388 , year =

    Qwen3 Technical Report , author =. arXiv preprint arXiv:2505.09388 , year =. doi:10.48550/arXiv.2505.09388 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.09388