Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

This paper claims that value alignment in multi-agent LLM systems is a system-level property, shaped by interaction topology and value type, and not reducible to individual agent alignment.

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

2026-08-03 03:13 UTC pith:FSCQXLQK

load-bearing objection Useful measurement framework for value drift in multi-agent LLMs, but the single-judge measurement layer and lack of variance reporting make the quantitative claims provisional. the 5 major comments →

arxiv 2602.08567 v2 pith:FSCQXLQK submitted 2026-02-09 cs.MA cs.CL

ValueFlow: Measuring the Propagation of Value Perturbations in Multi-Agent LLM Systems

Jinnuo Liu , Chuke Liu , Hua Shen This is my paper
classification cs.MA cs.CL
keywords value alignmentmulti-agent LLM systemsvalue drift propagationperturbation-based evaluationagent-level susceptibilitysystem susceptibilitynetwork topologyLLM-as-a-judge
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.

The paper tries to establish that how strongly a value perturbation propagates through a multi-agent LLM system is determined by the interplay of agent-level sensitivity and network structure, not by individual alignment alone. It introduces a perturbation-based measurement framework: agents' value orientations are scored on 56 value dimensions while they interact, small value-biased inputs are injected at chosen nodes, and two metrics track the response—one at the level of a single agent and one at the level of the system's outputs. Experiments across models, personas, values, and topologies show that widely shared normative values are hard to nudge, context-dependent values are easy to nudge, and central high-reachability positions amplify drift while many aggregating inputs dilute it. A sympathetic reader would care because the work provides a concrete way to audit which parts of a deployed multi-agent system are most at risk of value drift and where to intervene.

Core claim

The central claim is that value drift in multi-agent LLM systems is an emergent system-level property: for a fixed set of agents, whether a small value perturbation cascades into a system-wide shift depends on the type of value being probed and on the wiring of the interaction graph. The paper operationalizes this with β-susceptibility, the slope of an agent's output value score against the average value score of its peers' inputs, and system susceptibility (SS), the normalized average deviation of output nodes after a unit perturbation. Empirically, agent-level β predicts how far a perturbation travels through a fixed topology, and topology itself modulates propagation: reachability and cen

What carries the argument

The load-bearing machinery is a perturbation-and-measure loop. A fixed set of 56 values is probed by ten Yes–No questions per value; each agent's answers are scored 0–10 by a fixed LLM judge, with negative items inverted, so each invocation gets a value score. Perturbations are injected as synthetic peer responses optimized to push a target value to an extreme, and responses are fit to a linear model y = βx̄ + c, whose slope β is agent-level susceptibility; system susceptibility SS averages normalized output deviations over output nodes. The framework time-unrolls multi-round interaction into a static directed acyclic graph so that topology and agent behavior can be varied independently.

Load-bearing premise

The results rest on the assumption that the single fixed LLM judge's 0–10 scores are valid, value-discriminating measures of the agents' expressed value orientations; if this judge is biased or insensitive, the measured β and SS differences across values and topologies could be artifacts.

What would settle it

Run the β-susceptibility measurement for the 56 values under two or more independent LLM judges (or a re-scored human judge) and check whether the value-wise ranking of β and the topology effects in SS replicate; if the ranking changes materially across judges, the core empirical findings are measurement artifacts rather than system properties.

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

If this is right

  • Value-specific defenses become practical: values with consistently high β should be prioritized for monitoring, and prompt-level controls will have little effect on values that are stable under persona changes.
  • Topology-aware design emerges as a structural defense: decentralized interaction graphs with aggregation at sensitive nodes can attenuate value drift without modifying agent weights.
  • Evaluation practice should shift from isolated-model scores to paired agent-level and system-level susceptibility scores for multi-agent deployments.
  • The empirical portrait of values is a testable generalization: normative, widely shared values are intrinsically more resistant to peer influence than context-dependent ones across different backbones.

Where Pith is reading between the lines

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

  • The same perturbation-probing machinery could be repurposed to measure propagation of factual or stylistic biases in multi-agent systems, since the metric is agnostic to what the judge scores.
  • The strong dependence of β on input variance suggests a testable prediction beyond the paper: in real deployments, heterogeneous agent contexts may make systems more susceptible to value drift than homogeneous ones.
  • Because β is protocol-dependent, the specific numeric susceptibilities are unlikely to transfer across prompt or judge changes; the ordinal ranking of values and the structural effects are the more durable findings.
  • Replacing the single judge with a panel of independent judges would reveal which values show inflated apparent susceptibility due to measurement noise rather than genuine propagation.

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

5 major / 5 minor

Summary. The paper introduces ValueFlow, a framework for measuring how value perturbations propagate through multi-agent LLM systems. It constructs a 56-value question dataset derived from the Schwartz Value Survey, scores agent responses with a single fixed LLM-as-a-judge, and proposes two metrics: agent-level beta-susceptibility (slope of a linear fit between aggregated input value score and output score) and system susceptibility (SS, normalized output deviation under node perturbation). Experiments vary values, backbones, personas, input variance, and topology, leading to six findings, the central one being that value alignment in multi-agent systems is a system-level property shaped by interaction topology and value type, not just agent-level behavior. The paper includes a Limitations section acknowledging key weaknesses: the judge is unvalidated, beta is protocol-dependent, and only DAGs/time-unrolled graphs are considered.

Significance. If the measurement layer were validated, ValueFlow would provide a useful decomposition of value drift into agent-level and system-level components, and the topology findings (centrality, reachability, in-degree dilution) would be of practical interest for auditing multi-agent systems. The framework is clearly specified and reproducible in principle, and the paper ships a fixed dataset and perturbation prompts. However, the empirical support for the headline claims rests on a single LLM judge whose numeric scores are not shown to track human value judgments, and on an approximate linearity assumption with acknowledged low-R^2 cases. The structural findings (Finding 6) are more independent and are the most robust contribution; the relation between beta and SS (Finding 5) is, by the paper's own model, largely a consequence of linearity rather than a new empirical discovery.

major comments (5)
  1. [§2.2, Eq. (1), Appendix D.2] The entire quantitative layer depends on the LLM-as-a-judge scoring protocol, but the paper validates only question polarity (86.73% human agreement), not the validity of the 0-10 numeric scores. The few-shot examples in Table 5 are author-constructed and not human-calibrated. If the judge is biased—e.g., by answer length, verbosity, or a tendency toward mid-range scores—then beta and SS differences across values, personas, and backbones could be measurement artifacts. The Limitations paragraph acknowledges 'we do not study robustness across different evaluators or inter-judge variability,' but this is a load-bearing gap because all findings are computed from these scores. I ask for either a human rating study on a sample of responses or a multi-judge agreement analysis demonstrating that relative susceptibility rankings are stable.
  2. [§2.4.1, Eq. (3), Appendix E.1] The linear model in Eq. (3) is justified by 'approximately linear' response curves, but the manuscript also states that 'some value dimensions exhibit lower R2 values' and dismisses these as 'typically associated with small beta estimates.' No R2 values or residuals are reported. Low R2 in cases with small beta makes beta statistically unreliable, yet these beta values are used in cross-value comparisons (Finding 1) and across backbones (Table 1). I request that the paper report R2 (or confidence intervals) for the beta fits, at least for the representative values and for the aggregate distributions, and that the sensitivity of Findings 1-4 to excluding low-R2 values be reported.
  3. [§4.1, Finding 5] Finding 5—that high-beta values propagate farther and decay more slowly in a fixed layered topology—is to a large extent entailed by the linear model and the definition of SS. In a layered cascade with linear response y = beta*x + c, the output perturbation after d layers scales as beta^d, so SS is approximately proportional to beta^depth. The simulation in Figure 6 thus confirms the model's arithmetic rather than providing an independent empirical test. The paper should explicitly acknowledge this and either reframe Finding 5 as a model-derived prediction or design a system-level experiment that can falsify the linear-cascade prediction (e.g., by varying aggregation functions or introducing nonlinearities).
  4. [§3.1 and Table 1] All experiments appear to use a single run per configuration (t=0), and no error bars, confidence intervals, or seed variations are reported. This is particularly concerning for Table 1, where mean beta differences across backbones are small (e.g., 0.4078 vs 0.3245) and for Figure 4, where persona effects are compared visually. Without repeated sampling or a measure of variance, the reader cannot assess whether the reported differences are significant or within noise. At minimum, the paper should provide multiple independent runs with means and standard deviations for the headline comparisons, or justify why LLM stochasticity does not affect the conclusions.
  5. [Eq. (4)] The definition of SS normalizes by 'the magnitude of the injected perturbation' (Delta_pert), but this quantity is not operationally defined in the main text. It is ambiguous whether Delta_pert is the target score difference (e.g., 10 or 0 relative to baseline), the actual average input value shift at the perturbed node, or something else. This matters for comparing SS across topologies and perturbation locations, because the effective perturbation strength can vary with node in-degree and context. Please specify how Delta_pert is computed and, if it is a constant, state so explicitly.
minor comments (5)
  1. [Abstract and Introduction] The abstract alternates between '56-value dataset' and '56 human value dimensions'; consistent terminology would improve clarity.
  2. [Table 1 and Appendix F] The backbone names are inconsistently spelled: 'LLama-3.3-70B' in §3.1 but 'Llama-3.3-70B' in Appendix F. Please unify.
  3. [Figure 6] The color/symbol mapping for the three representative values (high/medium/low beta) is not described in the caption; please add a legend or textual explanation.
  4. [Appendix B] The human evaluation sampled 3 questions per value, but the dataset uses 7 positive and 3 negative questions per value. The polarity agreement of 86.73% is reported on a subset; please clarify the per-polarity agreement (positive vs. negative framing) to ensure negative questions are not systematically mislabeled.
  5. [Limitations] The Limitations section is candid and well-written, but its statement that 'evaluator-specific biases may still affect absolute scores' understates the risk for relative comparisons, since all findings derive from the same evaluator. This point should be moved to the main text or at least emphasized more strongly.

Circularity Check

1 steps flagged

Finding 5 largely restates the fitted β under the same perturbation protocol, but topology finding (Finding 6) gives independent content.

specific steps
  1. fitted input called prediction [Section 2.4.1 (Eq. 3), Section 2.4.2 (Eq. 4), Section 4.1 (Finding 5)]
    "We empirically observe an approximately linear relationship between yi and ¯xi across perturbation configurations. Therefore, we define β-susceptibility by fitting a linear model yi =β¯xi +c+ϵ i,(3) ... Finding 5: Under fixed topology, agent-level susceptibility strongly predicts system-level propagation dynamics: high-β values propagate farther and decay more slowly, while low-β values are rapidly corrected."

    β is the OLS slope of an agent's output score against the mean peer-input score in a controlled perturbation protocol. In the Section 4.1 layered fully-connected system, each layer's input is the previous layer's outputs, so under Eq. (3) the deviation at depth d scales as β^d; dividing by Δ_pert in Eq. (4) makes SS ≈ β^d. Hence 'high-β propagates farther' is an algebraic consequence of the fitted slope, not an independent empirical discovery. The finding is presented as a prediction but is forced by reusing the same perturbation protocol and the same linear assumption.

full rationale

ValueFlow's core quantitative layer is the LLM-as-a-judge scoring protocol; concerns about judge validity are correctness risks, not circularity, and the paper explicitly flags inter-evaluator robustness as a limitation. The only step approaching circularity is Finding 5: because β is fitted as the slope of output vs. mean peer input (Eq. 3) and SS normalizes output deviations by the injected perturbation (Eq. 4), the observed beta-to-SS relationship in a layered cascade is largely a restatement of the fitted linear response under the same protocol. This is a partial fitted-input-as-prediction reduction. However, the central claim that value alignment is a system-level property is not solely dependent on that step: Finding 6 compares SS across topologies and perturbation locations with agent behavior fixed, a contrast not derivable from Eq. (3) alone. No self-citation is load-bearing; prior author work is cited only as inspiration/context. Overall score reflects one moderate circular step rather than a fully circular derivation.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The framework is measurement, not derivation; the central claims depend on the validity of the LLM-judge, the linearity assumption, the DAG abstraction, and several hand-chosen constants (10 questions/value, threshold 6, perturbation prompt optimization). These are mostly reasonable design choices, but the judge validity is load-bearing and explicitly unvalidated.

free parameters (4)
  • beta_k (per-value agent-level susceptibility) = reported in Figures 3-4 and Table 1 (e.g., 0.60 for Gemma3-27B mean)
    Fitted via OLS in Eq. (3); the paper's main findings are rank-orderings of beta across values and models. Central measured quantity.
  • c_k (intercept in Eq. 3) = not reported
    Intercept in the linear model yi = beta*xbar_i + c + eps_i; required to estimate slope.
  • Perturbation direction threshold (baseline score 6) = 6.0
    Hand-chosen rule in Appendix C: if baseline score < 6, apply endorsement perturbation; otherwise rejection. Affects which perturbation is used and thus measured beta.
  • Number of questions per value (10) = 10
    Hand-chosen dataset size in Appendix B; value score is the mean over 10 questions, so this determines noise level of the measured scores.
axioms (5)
  • domain assumption Schwartz Value Survey's 56 values are a valid taxonomy for human values.
    Invoked in Section 2.2 as the basis for the evaluation dataset. If the taxonomy is not adequate for LLM value measurement, the framework's scope is limited.
  • ad hoc to paper LLM-as-a-judge yields valid value orientation scores.
    Invoked in Eq. (1) and Section D.2. The paper acknowledges in Limitations that evaluator biases may affect scores and robustness is not studied.
  • ad hoc to paper The relationship between aggregated input value score and output score is linear.
    Eq. (3) assumes linearity; Appendix E.1 argues 'approximately linear' but notes low R^2 cases. If nonlinearity is strong, beta is a poor summary.
  • domain assumption Multi-agent interaction can be represented as a time-unrolled DAG.
    Section 2.1 formalizes systems as DAGs. Limitations state that feedback loops and persistent memory are not covered.
  • ad hoc to paper COPRO-optimized perturbation prompts induce comparable extreme responses across values and backbones.
    Appendix C asserts 'comparable perturbation strength' without providing validation data. If prompts are not equally effective, beta comparisons across values may be confounded.

pith-pipeline@v1.3.0-alltime-deepseek · 16467 in / 13105 out tokens · 143895 ms · 2026-08-03T03:13:10.367431+00:00 · methodology

0 comments
read the original abstract

Multi-agent large language model (LLM) systems increasingly consist of agents that observe and respond to one another's outputs. While value alignment is typically evaluated for isolated models, how value perturbations propagate through agent interactions remains poorly understood. We present ValueFlow, a perturbation-based framework that measures value drift in multi-agent systems via a 56-value valuation dataset derived from the Schwartz Value Survey, with agent value orientations scored using an LLM-as-a-judge protocol. ValueFlow decomposes value drift into agent-level response behavior and system-level structural effects, captured by two metrics: \b{eta}-susceptibility, an agent's sensitivity to perturbed peer value signals, and system susceptibility (SS), the effect of node-level perturbations on final system outputs.Experiments span across value dimensions, backbones, personas, and topologies, showing that susceptibility varies sharply across values and is strongly shaped by interaction structure, indicating that value alignment in multi-agent systems is a system-level property, not just an agent-level one. ValueFlow thus provides a principled basis for auditing and mitigating value propagation in deployed multi-agent systems.

Figures

Figures reproduced from arXiv: 2602.08567 by Chuke Liu, Hua Shen, Jinnuo Liu.

Figure 1
Figure 1. Figure 1: Illustrative examples of value perturbation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the ValueFlow framework. The framework (i) models multi-agent interactions and quantifies agent-level value orientations; (ii) introduces controlled value perturbations; and (iii) measures value propagation using two metrics: agent-level susceptibility (β) and system-level susceptibility (SS). structure. Different values may therefore exhibit distinct propagation patterns, which cannot be cap￾t… view at source ↗
Figure 3
Figure 3. Figure 3: Value-wise agent-level β-susceptibility under a fixed agent configuration (Qwen3-8B, neutral openness persona). Values are sorted by their β scores. The distribution reveals substantial variation across value dimension. included in the input context of agent vi . Agent vi generates a response conditioned on the task query and the responses of its in-neighbors N− (vi). Agents are treated as black-box condit… view at source ↗
Figure 4
Figure 4. Figure 4: Agent-level β-susceptibility across value dimensions under different openness persona prompts (Qwen3- 8B). Each bar corresponds to one value and three colors show β under high-, neutral-, and low-openness personas. While high-openness generally increases susceptibility, the magnitude of this effect varies across values. orientation in response to peer signals under fixed interaction conditions. We focus on… view at source ↗
Figure 5
Figure 5. Figure 5: Distribution of agent-level β-susceptibility under high and low input context variance (Qwen3-8B, neutral openness persona). Each box summarizes β values across all 56 value dimensions. variance (low vs. high). Unless otherwise speci￾fied, Qwen3-8B, neutral openness, and high input variance are used as defaults. All other variables are held constant. 3.2 Variation Across Value Dimensions We first examine h… view at source ↗
Figure 6
Figure 6. Figure 6: System-level value propagation under different agent-level susceptibility regimes in a fully-connected topology. (A) Experimental setting: a fully-connected multi-agent system where all agents in the first round receives value perturbations, and remaining layers follow the original protocol. (B) System susceptibility (S S ) as a function of interaction depth for three representative values with high, mediu… view at source ↗
Figure 7
Figure 7. Figure 7: System-level value propagation under different network topologies and perturbation locations. Left: canonical interaction topologies considered. Right: system susceptibility (S S ) under single-node perturbations at different node positions. Higher reachability and structural centrality increase S S , while high in-degree at the perturbed node attenuates propagation. baseline within early rounds, even damp… view at source ↗
Figure 8
Figure 8. Figure 8: Examples of agent-level response curves un [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The detailed bar-plot for four backbone models’ 56-value [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Pseudo-Deliberation in Language Models: When Reasoning Fails to Align Values and Actions

    cs.CL 2026-05 unverdicted novelty 6.0

    LLMs exhibit pseudo-deliberation, with consistent value-action misalignment in generated dialogues despite reasoning, as measured by the new VALDI framework across 4941 scenarios.

  2. Pseudo-Deliberation in Language Models: When Reasoning Fails to Align Values and Actions

    cs.CL 2026-05 unverdicted novelty 6.0

    LLMs exhibit Pseudo-Deliberation where explicit reasoning fails to align stated values with generated actions, measured via the new VALDI framework across 4,941 scenarios in five domains.

Reference graph

Works this paper leans on

32 extracted references · 14 linked inside Pith · cited by 1 Pith paper

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Shuaihang Chen, Yuanxing Liu, Wei Han, Weinan Zhang, and Ting Liu. 2025. https://arxiv.org/abs/2412.17481 A survey on llm-based multi-agent system: Recent advances and new frontiers in application . Preprint, arXiv:2412.17481

    Pith/arXiv arXiv 2025

  4. [4]

    Christian Haerpfer, Ronald Inglehart, Alejandro Moreno, Christian Welzel, K Kizilova, Jaime Diez-Medrano, Marta Lagos, Pippa Norris, Eduard Ponarin, Bi Puranen, and 1 others. 2020. World values survey: Round seven-country-pooled datafile. madrid, spain & vienna, austria: Jd systems institute & wvsa secretariat. Version: http://www. worldvaluessurvey. org/...

    2020

  5. [5]

    Mengkang Hu, Pu Zhao, Can Xu, Qingfeng Sun, Jianguang Lou, Qingwei Lin, Ping Luo, and Saravan Rajmohan. 2025. https://arxiv.org/abs/2408.00764 Agentgen: Enhancing planning abilities for large language model based agent via environment and task generation . Preprint, arXiv:2408.00764

    Pith/arXiv arXiv 2025

  6. [6]

    Yoichi Ishibashi and Yoshimasa Nishimura. 2024. https://arxiv.org/abs/2404.02183 Self-organized agents: A llm multi-agent framework toward ultra large-scale code generation and optimization . Preprint, arXiv:2404.02183

    Pith/arXiv arXiv 2024

  7. [7]

    Han Jiang, Xiaoyuan Yi, Zhihua Wei, Ziang Xiao, Shu Wang, and Xing Xie. 2024 a . Raising the bar: Investigating the values of large language models via generative evolving testing. arXiv preprint arXiv:2406.14230

    Pith/arXiv arXiv 2024

  8. [8]

    Liwei Jiang, Sydney Levine, and Yejin Choi. 2024 b . https://openreview.net/forum?id=VUq1dDJBf0 Can language models reason about individualistic human values and preferences? In Pluralistic Alignment Workshop at NeurIPS 2024

    2024

  9. [9]

    Liwei Jiang, Taylor Sorensen, Sydney Levine, and Yejin Choi. 2025. https://arxiv.org/abs/2410.03868 Can language models reason about individualistic human values and preferences? Preprint, arXiv:2410.03868

    Pith/arXiv arXiv 2025

  10. [10]

    Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts

    Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. 2023. https://arxiv.org/abs/2310.03714 Dspy: Compiling declarative language model calls into self-improving pipelines . Preprint, arXiv:2310.03714

    Pith/arXiv arXiv 2023

  11. [11]

    Hannah Rose Kirk, Bertie Vidgen, Paul R \"o ttger, and Scott A Hale. 2024. The benefits, risks and bounds of personalizing the alignment of large language models to individuals. Nature Machine Intelligence, pages 1--10

    2024

  12. [12]

    Yunxuan Li, Yibing Du, Jiageng Zhang, Le Hou, Peter Grabowski, Yeqing Li, and Eugene Ie. 2024. https://arxiv.org/abs/2406.11776 Improving multi-agent debate with sparse communication topology . Preprint, arXiv:2406.11776

    Pith/arXiv arXiv 2024

  13. [13]

    Siyang Liu, Trisha Maturi, Bowen Yi, Siqi Shen, and Rada Mihalcea. 2024. The generation gap: Exploring age bias in the value systems of large language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 19617--19634

    2024

  14. [14]

    Yuanyi Ren, Haoran Ye, Hanjun Fang, Xin Zhang, and Guojie Song. 2024. https://arxiv.org/abs/2406.04214 Valuebench: Towards comprehensively evaluating value orientations and understanding of large language models . Preprint, arXiv:2406.04214

    Pith/arXiv arXiv 2024

  15. [15]

    Shalom H Schwartz. 1992. Universals in the content and structure of values: Theoretical advances and empirical tests in 20 countries. In Advances in experimental social psychology, volume 25, pages 1--65. Elsevier

    1992

  16. [16]

    Shalom H Schwartz. 1994. Are there universal aspects in the structure and contents of human values? Journal of social issues, 50(4):19--45

    1994

  17. [17]

    Shalom H Schwartz. 2012. An overview of the schwartz theory of basic values. Online readings in Psychology and Culture, 2(1):11

    2012

  18. [18]

    Shalom H Schwartz, Jan Cieciuch, Michele Vecchione, Eldad Davidov, Ronald Fischer, Constanze Beierlein, Alice Ramos, Markku Verkasalo, Jan-Erik L \"o nnqvist, Kursad Demirutku, and 1 others. 2012. Refining the theory of basic individual values. Journal of personality and social psychology, 103(4):663

    2012

  19. [19]

    Hua Shen, Nicholas Clark, and Tanushree Mitra. 2025. https://arxiv.org/abs/2501.15463 Mind the value-action gap: Do llms act in alignment with their values? Preprint, arXiv:2501.15463

    arXiv 2025

  20. [20]

    Hua Shen, Chieh-Yang Huang, Tongshuang Wu, and Ting-Hao Kenneth Huang. 2023. Convxai: Delivering heterogeneous ai explanations via conversations to support human-ai scientific writing. In Companion Publication of the 2023 Conference on Computer Supported Cooperative Work and Social Computing, pages 384--387

    2023

  21. [21]

    Hua Shen, Tiffany Knearem, Reshmi Ghosh, Kenan Alkiek, Kundan Krishna, Yachuan Liu, Ziqiao Ma, Savvas Petridis, Yi-Hao Peng, Li Qiwei, and 1 others. 2024 a . Towards bidirectional human-ai alignment: A systematic review for clarifications, framework, and future directions. arXiv preprint arXiv:2406.09264

    arXiv 2024

  22. [22]

    Hua Shen, Tiffany Knearem, Reshmi Ghosh, Yu-Ju Yang, Tanushree Mitra, and Yun Huang. 2024 b . Valuecompass: A framework of fundamental values for human-ai alignment. arXiv preprint arXiv:2409.09586

    arXiv 2024

  23. [23]

    Hua Shen, Yuguang Yang, Guoli Sun, Ryan Langman, Eunjung Han, Jasha Droppo, and Andreas Stolcke. 2022. Improving fairness in speaker verification via group-adapted fusion network. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7077--7081. IEEE

    2022

  24. [24]

    Taylor Sorensen, Jared Moore, Jillian Fisher, Mitchell Gordon, Niloofar Mireshghallah, Christopher Michael Rytting, Andre Ye, Liwei Jiang, Ximing Lu, Nouha Dziri, and 1 others. 2024. A roadmap to pluralistic alignment. arXiv:2402.05070

    Pith/arXiv arXiv 2024

  25. [25]

    Qiaosi Wang, Michael Madaio, Shaun Kane, Shivani Kapania, Michael Terry, and Lauren Wilcox. 2023. Designing responsible ai: Adaptations of ux practice to meet responsible ai challenges. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1--16

    2023

  26. [26]

    Zhenhailong Wang, Shaoguang Mao, Wenshan Wu, Tao Ge, Furu Wei, and Heng Ji. 2024. https://arxiv.org/abs/2307.05300 Unleashing the emergent cognitive synergy in large language models: A task-solving agent through multi-persona self-collaboration . Preprint, arXiv:2307.05300

    Pith/arXiv arXiv 2024

  27. [27]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. https://arxiv.org/abs/2201.11903 Chain-of-thought prompting elicits reasoning in large language models . Preprint, arXiv:2201.11903

    Pith/arXiv arXiv 2023

  28. [28]

    Zihao Yi, Jiarui Ouyang, Zhe Xu, Yuwen Liu, Tianhao Liao, Haohao Luo, and Ying Shen. 2025. https://arxiv.org/abs/2402.18013 A survey on recent advances in llm-based multi-turn dialogue systems . Preprint, arXiv:2402.18013

    Pith/arXiv arXiv 2025

  29. [29]

    Kechi Zhang, Jia Li, Ge Li, Xianjie Shi, and Zhi Jin. 2024. https://arxiv.org/abs/2401.07339 Codeagent: Enhancing code generation with tool-integrated agent systems for real-world repo-level coding challenges . Preprint, arXiv:2401.07339

    Pith/arXiv arXiv 2024

  30. [30]

    Xinyang Zhang, Ningfei Wang, Hua Shen, Shouling Ji, Xiapu Luo, and Ting Wang. 2020. Interpretable deep learning under fire. In 29th \ USENIX \ security symposium ( \ USENIX \ security 20)

    2020

  31. [31]

    Han Zhou, Xingchen Wan, Ruoxi Sun, Hamid Palangi, Shariq Iqbal, Ivan Vulić, Anna Korhonen, and Sercan Ö. Arık. 2025. https://arxiv.org/abs/2502.02533 Multi-agent design: Optimizing agents with better prompts and topologies . Preprint, arXiv:2502.02533

    arXiv 2025

  32. [32]

    Mingchen Zhuge, Wenyi Wang, Louis Kirsch, Francesco Faccio, Dmitrii Khizbullin, and Jürgen Schmidhuber. 2024. https://arxiv.org/abs/2402.16823 Language agents as optimizable graphs . Preprint, arXiv:2402.16823

    Pith/arXiv arXiv 2024