Pith. sign in

REVIEW 3 major objections 4 minor 3 cited by

Understanding How Value Neurons Shape the Generation of Specified Values in LLMs

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A 3% activation-probability gap between opposing value prompts marks the neurons that steer an LLM's value orientation.

desk verdict A useful dataset and neuron-localization study whose causal claim rests on an unvalidated GPT-4 judge; worth refereeing, but the central evidence needs human validation. read the letter →

arxiv 2505.17712 v1 pith:EMM35O5F submitted 2025-05-23 cs.CL

Yi Su , Jiayi Zhang , Shu Yang , Xinhai Wang , Lijie Hu , Di Wang This is my paper
classification cs.CL
keywords mechanisticinterpretabilityvaluealignmentSchwartzValuesSurveyneuronlocalizationactivationcontrasteditingG-EVALLLM
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper's goal is to show that the value orientation of a large language model is carried by a small set of identifiable neurons, and that those neurons can be located cheaply by comparing activations on opposing value descriptions. To do this, the authors build ValueInsight, a dataset of value descriptions and situational questions rooted in the Schwartz Values Survey, and define a neuron as value-related when its activation probability differs by more than 3% between a value and its reversed counterpart. They then amplify or suppress those neurons and show that G-EVAL scores move in the intended value direction, while the same manipulation of random neurons does not. If the claim holds, value alignment becomes a targeted, inspectable operation rather than a global retraining process.

What carries the argument

ValueLocate's load-bearing object is the differential activation probability. For a neuron in a given layer, the activation is accumulated over all token positions in a prompt; across many prompts, the activation probability is the empirical chance that the neuron activates at all. The contrast, written as $\delta = p^{+} - p^{-}$, isolates value selectivity by subtracting activation probability under the positive value description from that under its reversed counterpart. Neurons with $\delta$ above $+3\%$ are labeled positive-value neurons, those below $-3\%$ negative-value neurons, and the rest neutral. Validation uses an editing rule that suppresses negative-associated neurons, amplifies positive-associated ones by multiplying activation by $1 + \delta \gamma$, and leaves neutral neurons unchanged, so $\gamma$ is the dial controlling how strongly the value is pushed.

What would settle it

Have human raters, or a forced-choice value-priority test, judge the outputs produced by positive versus negative neuron edits without knowing which edit was applied; if they cannot reliably sort outputs by intended value direction, the causal claim fails. A second check is to apply G-EVAL to outputs that are stylistically rewritten toward a value without involving the located neurons; if scores rise as much as they do after real edits, the metric is measuring style rather than value.

Watch

Extended reading notes

Core claim

The paper claims that in transformer LLMs, value orientation is concentrated in a sparse set of feedforward neurons—under 0.4% of all neurons—and that the same sparse pattern appears across four value dimensions and several model families. The identification signal is differential activation: for each neuron, the activation probability on prompts carrying a positive value description is subtracted from that on prompts carrying the reversed description, and the signed difference locates which pole the neuron supports. Manipulating those neurons with a scaling factor changes G-EVAL scores monotonically in the edited direction, whereas manipulating randomly chosen neurons leaves scores essentially unchanged. This is offered as causal evidence that the located neurons genuinely shape the values the model expresses.

Load-bearing premise

The load-bearing premise is that G-EVAL scores reflect genuine value orientation and not merely stylistic alignment with the value descriptions, since the paper does not check the judge against human value judgments.

Editorial extensions

If this is right

  • Editing fewer than 0.4% of feedforward neurons moves a model's expressed value orientation, so value steering does not require retraining or full-model intervention.
  • The contrast method needs only paired opposing prompts, making neuron localization for values substantially cheaper than gradient-based attribution.
  • A scaling factor gives a monotonic control dial for value strength, letting one adjust how far the model's orientation is pushed.
  • The effect transfers across four model families and three evaluation datasets, suggesting the identified neurons are not artifacts of a single architecture or benchmark.
  • Since random neuron edits do not shift value scores, the located subset is functionally distinct from the surrounding network.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The core causal evidence inherits the limitations of G-EVAL: it is an automated judge scoring relevance to the same Schwartz dimensions used to author the value descriptions, so the paper does not rule out that the judge is detecting stylistic alignment rather than a genuine value change.
  • The same activation-contrast recipe could map other abstract attributes—ideology, personality, safety stance—by replacing Schwartz value descriptions with opposing description pairs for those attributes.
  • The 3% threshold is chosen from an ablation on one model; porting the method to other scales may require an adaptive threshold based on the distribution of $\delta$.
  • The observed middle-layer concentration suggests targeted layer-specific editing as a way to change values while disturbing other capabilities less.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes ValueLocate, a mechanistic interpretability framework for identifying neurons in LLMs that encode Schwartz value dimensions. It introduces ValueInsight, a dataset of 640 value descriptions and 15,000 scenario-based questions, and identifies candidate value-related neurons by the activation-probability difference (δ, Eq. 5) between positive and negative descriptions of the same value dimension. To validate causality, the paper edits these neurons by amplifying or suppressing their activations (Eq. 6) and measures the change in value orientation using G-EVAL, a GPT-4-based judge scoring responses on the same Schwartz dimensions. Experiments on four LLMs and three datasets report that ValueLocate outperforms baselines such as LPIP, QRNCA, and CGVST, and that editing identified neurons shifts G-EVAL scores in the expected direction while editing random neurons does not.

Significance. If the central claim holds, this is a valuable step toward mechanistic interpretability of value representations in LLMs, with practical implications for alignment. The paper has clear strengths: it contributes a new generative benchmark (ValueInsight) grounded in a psychological framework; the activation-contrast method is computationally cheap compared with attribution methods; it compares against multiple established neuron-localization baselines; it evaluates on four LLMs of varying scale; and it includes a random-neuron control that demonstrates specificity relative to random editing. The finding that only a small fraction (<0.4%) of neurons are value-related, concentrated in middle layers, is an interesting empirical observation. However, the validity of the causal claim rests entirely on G-EVAL scores, which are not validated against human judgment and share the same Schwartz taxonomy as the neuron-selection prompts.

major comments (3)
  1. [§5, §6.1, Eq. (6), Tables 1–6] The central causal claim is supported only by G-EVAL, a GPT-4 judge that scores responses for relevance to the same four Schwartz value dimensions used to construct the positive/negative description prompts for neuron identification (Eq. 5). The intervention in Eq. (6) amplifies neurons that fire more under positive descriptions and suppresses neurons that fire more under negative descriptions, so the generated responses are pushed toward the distribution of the positive descriptions. A GPT-4 judge could therefore assign higher scores because the outputs contain surface-level stylistic markers typical of those descriptions (e.g., phrases such as 'I value autonomy' or 'less-traveled experiences') rather than because the model's value orientation changed in a psychologically meaningful sense. The paper reports no human validation of G-EVAL on these dimensions, no agreement statistics (e.g., correlation or Cohen's kappa), and no independent behavioral measure (e.g., forced-choice value judgments). The evaluations on PVQ40 (Table 2) and ValueBench (Table 3) do not resolve this issue because those datasets are also scored by the same unvalidated G-EVAL judge. This is a load-bearing gap: if G-EVAL does not track human judgments of value orientation, the observed score changes are real but the interpretation that these are 'value-related' neurons is not supported.
  2. [§4, Eq. (5) and §6.3] The activation-difference threshold δ=3% is selected post hoc from the ablation results in Section 6.3, which use the same G-EVAL scores and a GPT-4o text-quality score to choose the threshold. This creates a selection-evaluation loop: the threshold is chosen to keep value scores high while stabilizing text quality, and the same value scores are then used to verify the causal effect of the neurons selected by that threshold. A principled selection procedure, such as a hold-out validation split or an a priori criterion, is needed. Similarly, the dynamic scaling factor γ is set to 2.0 in all main experiments without a stated selection criterion; Section 6.3 reports a monotonic increase in G-EVAL scores with γ, but no trade-off (e.g., text quality degradation) is quantified for the choice γ=2.0. Because the outcome measure is the same one used to calibrate these parameters, the reported effect sizes may be inflated.
  3. [§5, Figure 5] The random-neuron control in Figure 5 (and Figures 13–15) shows that randomly selected neurons do not produce the same score shifts, which is a useful control. However, it does not rule out the alternative explanation that the identified neurons are correlated with the generic linguistic contrast between the positive and negative descriptions (e.g., a positive versus negative tone, or a set of stylistic features such as expressing personal autonomy versus preferring structure). To support the claim that these are value-specific neurons, an additional control is needed where neurons are identified using a contrast between positive and negative prompts for an unrelated, non-value dimension, and then edited to show that value scores do not change in a dimension-specific way. Without such a control, the evidence is consistent with the neurons encoding stylistic features of the descriptions rather than Schwartz values per se.
minor comments (4)
  1. [§6.1, Table 1] The metric name is consistently typeset as 'G-EV AL' (with a space) in the text and tables; it should be 'G-EVAL'.
  2. [§4, Eq. (4)–(5)] The notation p+_l,k and p-_l,k in Eq. (5) is not formally defined; the definition of p_l,k in Eq. (4) should be instantiated separately for prompts containing the positive value description and its reversed counterpart.
  3. [Abstract and §7] The abstract states that the method 'establish[es] causal relationships between neurons and value representations.' Given that the outcome measure is an unvalidated LLM judge and the selection parameters are calibrated on that judge, a more cautious claim (e.g., 'suggest a causal role') would be more accurate pending human validation.
  4. [§3, Figure 1] The description of ValueInsight construction does not report any manual or automatic quality checks for the 15,000 situational questions apart from a GPT-4o refinement pass; a brief description of sample sizes per atomic value and a human spot-check would strengthen the dataset credibility.

Circularity Check

2 steps flagged · score 6.0 of 10

The central validation loop is partially circular: the 3% neuron-selection threshold is fitted to the same G-EVAL value score later used as evidence, and the G-EVAL judge scores the same Schwartz dimensions that define the positive/negative contrast prompts used to select the neurons.

  1. fitted input called prediction [Section 4 (Value Related Neuron Identification) and Section 6.3 (Effect of the Difference Threshold), with Eq. (5)-(6) and Tables 1-6.]
    "We chose a value threshold of 3% as our experiments in Section 6.3 show that it marks the point where the value score remains relatively high while the text quality stabilizes. ... The results confirm our intuition, leading us to choose a threshold of 0.03, as it represents the point where text quality stabilizes while maintaining relatively high value scores."

    The 3% threshold in Eq. (5) determines which neurons are labeled value-related and are edited in Eq. (6). Section 6.3 selects this threshold by inspecting the G-EVAL value score on the same evaluation setup, and Tables 1-6 then use that same G-EVAL score to claim that ValueLocate outperforms baselines and that manipulation changes value orientation. The confirmation is therefore partly an in-sample fit: the parameter controlling the neuron set was chosen to keep the outcome metric high. Comparisons with baselines are also run at a threshold tuned to ValueLocate's neuron distribution rather than to each baseline, so the reported superiority is not an independent prediction. Random-neuron and direction controls retain some empirical content, which is why the circularity is partial.

  2. self definitional [Section 6.1 (Evaluation Metric) with Section 4 Eq. (5) and ValueInsight construction in Section 3.]
    "This score reflects the relevance to a specific value dimension in the Schwartz Values Survey, with higher scores indicating a stronger presence of that value. ... δ=p+l,k−p−l,k, where p+l,k and p−l,k denote the activation probability of neuron computed from prompts containing the target value description (positive value) and its reversed counterpart (negative value)."

    Value-related neurons are selected by contrasting positive and negative value descriptions built from Schwartz dimensions, while G-EVAL asks GPT-4 to rate the response's relevance to the same Schwartz dimension. The intervention amplifies the positive-description activation pattern and suppresses the negative one, so a judge keyed to that same taxonomy may assign higher scores because the output inherits the wording and tone of the positive descriptions, not because an independently validated value orientation changed. The PVQ40 and ValueBench checks are scored by the same unvalidated G-EVAL judge, so they do not break the loop. No human agreement statistics are reported for G-EVAL on these dimensions.

full rationale

The paper's non-circular components are substantial: random-neuron editing controls, directional reversal, multiple model families, and comparison against three neuron-localization baselines all provide evidence that the identified neurons are not arbitrary. No load-bearing self-citation chain or imported uniqueness theorem appears. However, the central causal claim is weakened by two construction-level loops. First, the 3% threshold that defines the value-neuron set is tuned on the G-EVAL value score, and the same score is then presented as the outcome confirming neuron relevance. Second, both sides of the validation use the same Schwartz taxonomy: the contrast prompts that generate δ and the GPT-4 judge's relevance rubric are operationalized from the same four value dimensions, with no human validation of the judge on this construct. These loops make the strongest claim---that the manipulation changes the model's value orientation in a psychologically meaningful sense---partly self-referential, although not fully forced. The appropriate score is therefore 6, not 8 or 10, because the random controls and cross-model/baseline comparisons still carry independent empirical content.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the validity of the Schwartz framework, the GPT-4o generated antonym descriptions, and the G-EVAL metric. The threshold delta and scaling gamma are free parameters chosen from ablations on the same evaluation setup.

free parameters (2)
  • activation difference threshold delta = 0.03
    Selected in Section 6.3 ablation as the point where text quality stabilizes while value scores remain high. This is fitted to the same models and evaluation metric used for the headline results.
  • dynamic scaling factor gamma = 2.0
    Set to 2.0 for all main experiments. Ablation shows a monotonic effect but no principled selection criterion; it controls the magnitude of neuron manipulation.
assumptions (4)
  • domain assumption The Schwartz Values Survey provides a valid taxonomy of universal human values.
    The entire dataset and evaluation are built on this framework; the paper assumes the four dimensions are meaningful and that opposing aspects are well-defined.
  • domain assumption GPT-4o generated antonym descriptions faithfully represent opposing value orientations.
    ValueInsight uses GPT-4o generated second-person narratives for positive and negative value aspects, with manual refinement; no independent validation that these are genuine opposites.
  • domain assumption G-EVAL (GPT-4 based) is a valid measure of value orientation.
    All outcome scores are produced by a GPT-4 judge; the paper does not validate G-EVAL against human judgments or other independent metrics.
  • ad hoc to paper Neuron activation probability contrast isolates value-related processing rather than generic prompt style differences.
    The identification method assumes that activation differences between positive and negative prompts are specific to values, not to other contextual features. The random-neuron control partially addresses this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Understanding How Value Neurons Shape the Generation of Specified Values in LLMs." pith.science (2026). https://pith.science/paper/EMM35O5F

@misc{pith2026250517712,
  author       = {Pith},
  title        = {Pith review of: Understanding How Value Neurons Shape the Generation of Specified Values in LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EMM35O5F}},
  note         = {Machine review of arXiv:2505.17712}
}
read the original abstract

Rapid integration of large language models (LLMs) into societal applications has intensified concerns about their alignment with universal ethical principles, as their internal value representations remain opaque despite behavioral alignment advancements. Current approaches struggle to systematically interpret how values are encoded in neural architectures, limited by datasets that prioritize superficial judgments over mechanistic analysis. We introduce ValueLocate, a mechanistic interpretability framework grounded in the Schwartz Values Survey, to address this gap. Our method first constructs ValueInsight, a dataset that operationalizes four dimensions of universal value through behavioral contexts in the real world. Leveraging this dataset, we develop a neuron identification method that calculates activation differences between opposing value aspects, enabling precise localization of value-critical neurons without relying on computationally intensive attribution methods. Our proposed validation method demonstrates that targeted manipulation of these neurons effectively alters model value orientations, establishing causal relationships between neurons and value representations. This work advances the foundation for value alignment by bridging psychological value frameworks with neuron analysis in LLMs.

Figures

Figures reproduced from arXiv: 2505.17712 by the authors.

Figure 1
Figure 1. ValueInsight Construction and Usage flect a wide variety of real-life behaviors. To fur￾ther enrich our set of questions, we incorporate common topics of life T from UltraChat (Ding et al., 2023), including family, environment, and arts. To generate these situational questions, we use specially formulated prompts P for GPT-4o. These prompts are designed to facilitate the gen￾eration of complex scenarios that involve… view at source ↗
Figure 2
Figure 2. Mainstream process of ValueLocate its reversed counterpart in ValueInsight, we prompt LLM to answer situational questions accordingly. During this process, we calculate the neuron acti￾vation value ml k for an input sequence x of length T: ml k = X T i=1 σ(Wl 1k · (h l−1 i + A l i )), (3) where Wl 1k is the k-th row of Wl 1 . Given N input sequences, each comprising a de￾scription and a corresponding situational que… view at source ↗
Figure 3
Figure 3. Results of positively and negatively editing the neurons identified by ValueLocate, as well as editing [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (10 more)
Figure 5
Figure 5. Figure 5: Impact of Value-Related Neuron and Random [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: How threshold influences the result on LLama [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 9
Figure 9. Figure 9: LLama-3.2-1B Neuron Distribution [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: how threshold influences the result on LLama-3.1-8B for Self-Transcendence [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 7
Figure 7. Figure 7: Qwen2-0.5B Neuron Distribution [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 12
Figure 12. Figure 12: how threshold influences the result on LLama-3.1-8B for Conservation 15 [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 13
Figure 13. Figure 13: Impact of Value-Related Neuron and Ran￾dom Neuron Manipulation on Qwen2-0.5B [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]
Figure 14
Figure 14. Figure 14: Impact of Value-Related Neuron and Ran￾dom Neuron Manipulation on LLama-3.2-1B [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 15
Figure 15. Figure 15: Impact of Value-Related Neuron and Ran￾dom Neuron Manipulation on gemma-2-9B 16 [PITH_FULL_IMAGE:figures/full_fig_p016_15.png]
Figure 16
Figure 16. Figure 16: Results of positively and negatively editing the neurons identified by ValueLocate, as well as editing [PITH_FULL_IMAGE:figures/full_fig_p020_16.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. Dual Mechanisms of Value Expression: Intrinsic vs. Prompted Values in Large Language Models

    cs.CL 2025-09 conditional novelty 6.0 of 10

    Intrinsic and prompted value expressions in LLMs share some circuits but have distinct mechanisms: intrinsic directions make outputs more diverse, prompted directions drive instruction compliance and can disable refusal.

  2. COMPKE: Complex Question Answering under Knowledge Editing

    cs.CL 2025-06 conditional novelty 6.0 of 10

    COMPKE is a new benchmark with 11,924 complex questions that tests knowledge editing through one-to-many relations and logical operations, where existing editing methods often fail.

  3. The Compositional Architecture of Regret in Large Language Models

    cs.CL 2025-06 reject novelty 5.0 of 10

    The paper claims that regret in LLMs is encoded by interacting neuron groups detectable in the final hidden layer, using new S-CDI, RDS, and GIC metrics.

Reference graph

Works this paper leans on

24 extracted references · 21 canonical work pages · cited by 3 Pith papers

  1. [2]

    Keyi Kong, Xilie Xu, Di Wang, Jingfeng Zhang, and Mohan S Kankanhalli

    Alignment of language agents.arXiv preprint arXiv:2103.14659. Keyi Kong, Xilie Xu, Di Wang, Jingfeng Zhang, and Mohan S Kankanhalli. 2024. Perplexity-aware cor- rection for robust alignment with noisy preferences. Advances in Neural Information Processing Systems, 37:28296–28321. Jiyoung Lee, Minwoo Kim, Seungho Kim, Junghwan Kim, Seunghyun Won, Hwaran Le...

    arXiv 2024

  2. [3]

    While the overall topic should align with the “TOPIC”, each question should explore a different subtopic and situation to avoid repetition

    work page

  3. [4]

    honest" or

    Avoid tendency qualifiers like "honest" or "polite"

    work page

  4. [5]

    Politeness (demon- strating courtesy and social etiquette),

    Provide questions directly, each on a new line, without additional explanation. B Introduction to Schwartz Value Survey Developed through rigorous cross-cultural valida- tion studies, the Schwartz Value Survey constitutes a psychometric instrument comprising 56 items that operationalize 11 fundamental motivational domains: Achievement, Benevolence, Confor...

    work page

  5. [6]

    generate value descriptions Given a set of keywords associated with the value {value}: {key}, {key}, {key}

    avoid containing the key words or values in the descriptions. generate value descriptions Given a set of keywords associated with the value {value}: {key}, {key}, {key}.... along with an example second-person de- scription reflecting value , and an exam- ple second-person description reflecting its antonyms. I want you to give 80 second- person views of t...

    work page

  6. [7]

    each view should be related to the key words

    work page

  7. [8]

    {factor}

    avoid containing the key words or values in the descriptions. ###Exam- ple:{value description} ###Example for its antonyms:{value description} 11 text quality evaluation Please act as an impartial judge and evalu- ate the text quality of the provided content. Focus only on whether the expression is clear and fluent, and whether there are any grammatical e...

    work page

  8. [9]

    Ensure the revised question includes a similar and specific scenario and remains relevant to the factor

    work page

Show all 24 references
  1. [10]

    honest",

    Avoid tendency qualifiers like "honest", "polite" and similar. 12 generate value situational questions I want you to create a set of 10 situational questions aimed at evaluating the degree to which the respondent displays the specified "V ALUE", referring to the "EXAMPLE". Ple...

  2. [11]

    Try your best to create detailed and complex scenarios of at least 100 words for each question, focusing on specific dilem- mas, conflicting priorities, or challenging choices

  3. [12]

    V ALUE" and strictly limit them to

    Ensure questions are directly related to the "V ALUE" and strictly limit them to "What do you think" and "What would you do"

  4. [16]

    Openness to Change: Self-Direction (Creativ- ity, Freedom, Independent, Curious, Choos- ing own goals), Stimulation (A varied life, An exciting life, Daring), Hedonism (Pleasure, Enjoying life)

  5. [17]

    Self-Transcendence: Universalism (Broad- mindedness, Wisdom, Social justice, Equal- ity, A world at peace, Protecting the environ- ment, Unity with nature, A world of beauty), Benevolence (Helpfulness, Honesty, Forgive- ness, Loyalty, Responsibility, True friendship, Mature love)

  6. [18]

    Conservation: Tradition (Respect for tradition, Humility, Devoutness, Moderation), Confor- mity (Self-discipline, Obedience, Politeness, Honoring of parents and elders), Security (Na- tional security, Family security, Social order, Cleanliness, Reciprocation of favors, Health,...

  7. [19]

    Not like me at all

    Self-Enhancement: Achievement (Success, Capability, Intelligence, Ambition, Influence), Power (Social power, Authority, Wealth, Preservation of one’s public image, Social recognition), Hedonism (Pleasure, Enjoying life). 13 C Introduction about evaluation datasets C.1 PVQ40 Th...

  8. [20]

    She likes to do things in her own original way

    Thinking up new ideas and being creative is important to her. She likes to do things in her own original way

  9. [21]

    She wants to have a lot of money and expensive things

    It is important to her to be rich. She wants to have a lot of money and expensive things

  10. [22]

    She believes everyone should have equal opportunities in life

    She thinks it is important that every person in the world be treated equally. She believes everyone should have equal opportunities in life

  11. [23]

    She wants people to admire what she does

    It’s very important to her to show her abilities. She wants people to admire what she does. C.2 ValueBench ValueBench is the first comprehensive psychomet- ric benchmark designed to evaluate value orienta- tions and value understanding in LLMs. It aggre- gates data from 44 est...

  12. [24]

    Value Descriptions: Definitions and hierarchi- cal relationships (e.g., Schwartz Values Survey)

  13. [25]

    query neurons

    Item-Value Pairs: 15,000+ expert-annotated linguistic expressions (items) linked to specific val- ues. D Introduction about baselines D.1 LPIP The LPIP (Log Probability and Inner Products) method is a static approach designed to identify critical neurons in LLMs that contribut...

  14. [2021]

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy

    A mathematical framework for transformer circuits.Transformer Circuits Thread, 1(1):12. Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer feed-forward layers are key-value memories. InProceedings of the 2021 Conference on Empirical Methods in Natural L...

    2021

  15. [2022]

    Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Mered- ith Ringel Morris, Percy Liang, and Michael S Bern- stein

    Training language models to follow instruc- tions with human feedback.Advances in neural in- formation processing systems, 35:27730–27744. Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Mered- ith Ringel Morris, Percy Liang, and Michael S Bern- stein. 2023. Generative agents:...

    2023 arXiv

  16. [2024]

    Advances in Neural Information Processing Systems, 36

    Evaluating the moral beliefs encoded in llms. Advances in Neural Information Processing Systems, 36. Shalom H Schwartz, Gila Melech, Arielle Lehmann, Steven Burgess, Mari Harris, and Vicki Owens. 2001. Extending the cross-cultural validity of the theory of basic human values w...

    2017 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.