Pith. sign in

REVIEW 4 major objections 4 minor 82 references

Requiring LLMs to sort input into facts, heuristics, and emotions before deciding measurably reduces repeated-output variability and decision flips across 26 models.

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 12:09 UTC pith:RGBDZ76C

load-bearing objection Worth a serious look: the single-turn SI effect is credible and well-controlled, but the headline 82% DFR reduction is not yet attributable to F/H/E because the persistence comparison lacks a vanilla-persistence baseline. the 4 major comments →

arxiv 2607.24765 v1 pith:RGBDZ76C submitted 2026-06-05 cs.CL cs.AIcs.HC

Measuring and Improving Behavioral Consistency in Large Language Models through Fact-Heuristic-Emotion State Enforcement

Gi-Hun Lee , Joong Yull Park This is my paper
classification cs.CL cs.AIcs.HC
keywords behavioral consistencydecision flip ratesemantic integrityfact-heuristic-emotion decompositionstate persistenceprompt-level interventionlarge language modelsKorean-language decision scenarios
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 targets a deployment problem: an LLM can answer the same decision prompt differently across runs, and can reverse its choice when its own previous answer is fed back as context. The proposed remedy is the Cognitive Kernel Model (CKM), a prompt-level layer that forces the model to sort each input into Fact (given/verifiable), Heuristic (inferred/assumed), and Emotion (evaluative/priority) before choosing an action. In Korean-language decision scenarios across 26 models, the intervention raised the semantic consistency of repeated answers (random-effects Hedges' g≈1.09) and, with state persistence, cut the decision-flip rate from 0.385 to 0.069 in newer models. Ablations suggest the gain is not JSON formatting: the semantic separation survives value-only embedding, while the format-only effect disappears. The paper explicitly limits the claim to stability—not improved reasoning correctness.

Core claim

The paper's central claim is that one form of LLM behavioral instability—variation in repeated responses and reversal of decisions when prior output is reinjected—is measurable and partly reducible without changing weights. The mechanism is an enforced epistemic ledger: before acting, the model must tag each piece of information as Fact (given/verifiable), Heuristic (inferred/assumed), or Emotion (evaluative/priority), and optionally carry that structured state forward. In the tested protocol this raised Semantic Integrity (SI) by roughly +0.02 while halving the temperature-induced consistency penalty, and reduced the Decision Flip Rate under state persistence by 82% (0.385→0.069) in the new

What carries the argument

The Cognitive Kernel Model (CKM) — a prompt-level state-enforcement layer that maintains S_t = {F_t, H_t, E_t}, the structured record of what is given, what is inferred, and what is valued. It forces role separation before decision, and the CKM_PERSIST variant injects the Turn-1 structured state as Turn-2 context. The workhorse is the epistemic distinction itself: token streams arrive undifferentiated, so externally labeling provenance reduces ambiguity in the reasoning state. Supporting machinery includes four discriminative metrics (SI, DFR, FCS, SDR) and the sham-restriction arms that separate structural scaffolding (~45%) from reasoning-semantic content (~55%).

Load-bearing premise

The load-bearing premise is that the drop in decision flips comes from the Fact/Heuristic/Emotion labels themselves and not from the mere act of feeding the model its own previous answer back as context—a control the main comparison does not include, and which the paper's own ablation table suggests may account for most of the remaining margin.

What would settle it

Run the same 12-model, 8-scenario protocol with a Vanilla-persistence arm: inject the model's own free-form Turn-1 answer as Turn-2 context, no F/H/E schema. If DFR drops from ~0.385 to ~0.069, the flip reduction is self-answer reinjection and the F/H/E labels contribute little; if DFR stays near 0.385, the epistemic role separation is causal.

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

If this is right

  • Deployment teams can profile candidate models with the four discriminative metrics (SI, DFR, FCS, SDR) before committing, because single-turn semantic stability does not predict multi-turn decision stability (Tier B models show high SI with high DFR).
  • State-persistence compatibility is a measurable, model-dependent property: DFR spans 0.01–0.85 in legacy models, and 6 of 12 newer models reach perfect decision preservation under CKM persistence.
  • Multi-turn instability in this setting is primarily a state-construction and injection problem, not sampling randomness: FCS≈1.0 under fixed anchor states.
  • A model-agnostic prompt-level layer can improve consistency across vendors and generations without fine-tuning, and the benefit grows under stochastic sampling (Hedges' g=2.87 at temperature 0.7).
  • The gain decomposes into roughly 45% structural scaffolding and 55% F/H/E semantic content; only the epistemic-content arm raises SI and lowers DFR simultaneously, so structurally rich non-epistemic schemas can even destabilize decisions (Sham-B DFR=0.419).

Where Pith is reading between the lines

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

  • Extension — A plain-context persistence arm (inject the model's own free-form answer without any schema) would isolate whether the 82% flip reduction comes from F/H/E labels or from the act of re-injecting one's own answer; the paper's existing 5-arm table already shows Vanilla persistence at DFR 0.080, close to CKM's 0.069.
  • Extension — The Legacy cohort's 2.4× larger CKM gain on ambiguity-heavy scenarios suggests the separation works by reducing fact-inference conflation; a natural test is to apply the layer to professional or safety-critical decisions with high stakes and mixed fact/inference content.
  • Extension — The Sham-B result, where non-epistemic contextual slots raised DFR to 0.419, predicts that agent memory systems need provenance tags, not just retention; injecting semantically rich but unanchored context can actively destabilize decisions.
  • Extension — If the mechanism is epistemic provenance labeling, the effect should transfer across languages and domains. Matched English/Korean/Japanese scenarios with expert quality ratings would separate 'more stable' from 'better', testing the paper's explicit non-claim about correctness.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes the Cognitive Kernel Model (CKM), a prompt-level intervention that forces an LLM to decompose decision inputs into Fact, Heuristic, and Emotion fields before answering, and optionally carries the structured state into a second turn. Using Korean student decision scenarios across 26 models, four vendors, and 37,403 observations, it reports that CKM raises Semantic Integrity (SI) relative to free-form control (combined random-effects g = 1.09, 31 model-level pairs), that the effect survives value-only and sham-restriction controls, and that state persistence lowers the decision-flip rate (DFR) by 82% in a 12-model NewModels cohort. The paper is careful to restrict claims to consistency rather than correctness, and it reports extensive robustness and ablation work, including a temperature probe, a five-arm sham experiment, and an excluded-model case study.

Significance. The SI finding is a credible empirical contribution: it is cross-vendor, reproducible (hashed canonical datasets, released code), and the authors constructed controls—value-only SI, FHE-Only, and a five-arm sham ablation—that address the most obvious formatting confound. The temperature probe and the transparent case study of gpt-4.1-mini add value. However, the paper's headline multi-turn claim (state persistence reduces flipping via F/H/E separation) is not yet supported because the persistence contrast lacks a vanilla self-injection control, and the Exp6 comparison misaligns protocols across arms. If a vanilla-persistence control is added or the claim is restricted to the single-turn SI effect, the paper would make a solid contribution to LLM consistency evaluation.

major comments (4)
  1. [§5.3 / §6.2 / Table 7] The headline DFR reduction (0.385 → 0.069, g = 1.52) compares the two-turn CKM_PERSIST condition (Turn 1 output injected before Turn 2) with the single-turn CKM condition. Since DFR by definition (§4.3) is disagreement between T1 and T2, this contrast measures the effect of injecting one's own prior answer, not specifically the effect of F/H/E state. A vanilla-persistence arm (free-form Turn 1 appended verbatim before Turn 2) is absent. Exp6's Vanilla DFR_persist = 0.080 is explicitly a single-turn repetition flip rate (Table 10 note), not a persistence baseline. Without the missing control, the 82% reduction may be generic self-anchoring. Please add the control or meaningfully downgrade the persistence claim.
  2. [§6.5, Table 10 / Table 13] The SI–DFR dissociation and the conclusion that only CKM improves both axes are confounded by protocol mismatch. The footnote to Table 10 states that Vanilla/Sham-A/B/C are single-turn while CKM's DFR_persist comes from the two-turn CKM_PERSIST condition. Thus ΔDFR in Table 13 compares different tasks across arms. The Sham-B 'paradox' may simply reflect a difference between single-turn repetition flip rates and two-turn state-injection flip rates, unrelated to slot semantics. Align all arms to the same persistence protocol before claiming that sham arms degrade DFR.
  3. [§5.2 (gpt-4.1-mini exclusion)] The primary NewModels cohort excludes gpt-4.1-mini using criteria (3) and (4), which the paper explicitly labels as outcome-dependent and operationalized after data inspection ('anomalous behavioral profile'; 'directional preservation'). The transparency and sensitivity analysis are commendable, but the headline '12/12 positive effects' and all primary N = 12 statistics are conditional on an exclusion rule that used outcome data. Please present the 13-model analysis as co-primary, or clearly label the 12-model results as a post-hoc sensitivity subset rather than the primary cohort.
  4. [§6.3, Figure 5] The 'behavioral-variance reducer' correlation (r = −0.653 between baseline SI and ΔSI) is expected under a ceiling effect: SI is bounded above by 1.0, and the Control baseline sits at 0.907–0.970, leaving more headroom for lower-baseline models. No ceiling correction or null model (e.g., regression to the mean under additive noise) is provided, so the claim that intrinsically less consistent models benefit more is not established. This finding is secondary, but it should be reanalyzed or explicitly described as a descriptive ceiling artifact.
minor comments (4)
  1. [§4.1 / §6.1] SI is defined on 'responses', but the primary analysis uses parsed-field text (fact list + reason). The raw-output basis gives g = 0.61 and is reported only in the supplement. Please state in each table which representation is embedded, and justify why parsed-field is not a selective reporting choice.
  2. [Figures 4 and 5] The SI axes are truncated (beginning at 0.92 and 0.90, respectively) without explicit truncated-axis markers. This visually exaggerates separation. Use zero-origin or clearly marked axis breaks.
  3. [§6.5] The phrase 'claim strength: below β' (and the 'Judgment: R4' framing) is unexplained jargon in the main text. Define the claim-strength scale and R4 scenario, or move the framing to the supplement.
  4. [Global] There are typographical artifacts such as 'sufficient' and inconsistent use of 'CKM_PERSIST' vs. 'persist'. A careful proofreading pass is needed.

Circularity Check

0 steps flagged

No circular derivation; empirical contrasts are self-contained, with a non-circular DFR control caveat.

full rationale

The paper's load-bearing claims are empirical contrasts, not derivations: SI gains are measured between Control and CKM/FHE-Only arms; DFR is measured across T1/T2 protocols; the ablation decompositions are arithmetic on observed means. No equation defines a predicted effect in terms of a fitted input. The only construction-like risk is SI: a fixed JSON slot template could mechanically raise cosine similarity. The paper directly controls this with parsed-field SI, value-only SI with keys/syntax stripped, and the FHE-Only free-form arm (SI 0.981 vs Control 0.960, Table 5), so the SI effect does not reduce to format by construction. The DFR state-persistence claim has a genuine control limitation: the Table 10 footnote states "single-turn for Vanilla / Sham-A/B/C; CKM_PERSIST for the CKM arm," so the "only arm that reduces DFR" comparison lacks a vanilla-persistence baseline, and the 82% reduction (Table 7: CKM 0.385 vs CKM_PERSIST 0.069) could reflect self-anchoring. But this is a causal-identification confound, not circularity: no parameter is fitted to the outcome and then renamed a prediction, and DFR improvement is not definitionally equal to the F/H/E intervention. There are no load-bearing self-citations, no imported uniqueness theorem (the paper explicitly disclaims uniqueness of F/H/E), and no ansatz smuggled via citation. Within the circularity rubric, the derivation chain is self-contained.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 3 invented entities

The central claims rest on empirical comparisons, not on a formal derivation. The main free choices are the metric thresholds, the embedding backbone, and the outcome-dependent cohort exclusion. The F/H/E state and its proposed extensions are invented conceptual objects whose only evidence is the paper's own experiments.

free parameters (3)
  • Tier cutoffs (SI, DFR) = SI 0.95/0.97; DFR 0.15
    Posterior cutoffs chosen to land near observed gaps in the SI–DFR joint distribution (§6.3); used for the descriptive tier taxonomy, not primary claims.
  • CRR similarity threshold θ = 0.75
    Chosen for the Context Retention Rate metric (§4.2); the metric is reported as non-discriminative, so it does not affect primary claims.
  • Embedding backbone selection = paraphrase-multilingual-mpnet-base-v2
    Selected after sensitivity analysis showed miniLM yielded null effect (d=0.05); the choice to adopt the backbone that produces the effect is a hand-selected config that affects the primary SI results (§4.1, §6.4).
axioms (4)
  • domain assumption Sentence-BERT cosine similarity is a valid proxy for semantic consistency.
    SI, SDR, and CRR all rely on embedding cosine; the paper's own sensitivity analysis shows the backbone choice flips results to null (§6.4, §S1).
  • domain assumption Model-reported F/H/E labels faithfully represent epistemic-role separation.
    24k+ LLM outputs are validated for schema compliance (98.4%), not for whether the content truly is Fact/Heuristic/Emotion; the human pilot covers only 15 CKM responses (§3.3, §7.3).
  • domain assumption Model-level means are a valid unit of analysis.
    Inferential tests treat each model as one paired observation (n=12–31), ignoring within-model and scenario variance; no mixed-effects model is fit, so small n inflates g and p-values (§5.4; the paper itself calls for mixed-effects models as future work in §7.6).
  • domain assumption API sampling at a single temperature with provider defaults is comparable across vendors.
    No seeds, top-p defaults, and gpt-5-mini at T=1.0; conclusions assume vendor-side nondeterminism is noise rather than a confound (Reproducibility Statement).
invented entities (3)
  • CKM structured state S_t = {F_t, H_t, E_t} no independent evidence
    purpose: Prompt-level epistemic-role separation ledger that constitutes the intervention; also the measurement substrate for SI/DFR/SDR.
    The state is a prompt schema whose effect is the measured outcome; there is no falsifiable handle outside the paper's own experiments, though code/prompts are shipped for reproducibility.
  • Tri-Engine architecture (Ψ-Architect, DecisionEngine, Meta-Structure) no independent evidence
    purpose: Proposed extension layer that operates on the normalized kernel state; not evaluated.
    Mentioned in §3.2 as future work; no empirical evidence is provided.
  • Testimony (T) provenance axis no independent evidence
    purpose: Fourth epistemic axis proposed for future work (§7.6); also covered by a pending patent application.
    Not evaluated; no falsifiable handle outside the paper; disclosed as an unvalidated extension.

pith-pipeline@v1.3.0-alltime-deepseek · 32829 in / 14955 out tokens · 155716 ms · 2026-08-02T12:09:56.838204+00:00 · methodology

0 comments
read the original abstract

Large language models (LLMs) can give different answers to the same decision problem across runs, and reverse a decision when their own prior answer returns as context. We ask whether this instability can be measured and partially reduced without changing model weights. We test the Cognitive Kernel Model (CKM), a prompt-level state-enforcement layer. Before deciding, the model must separate its input into three epistemic roles: Fact (given or verifiable), Heuristic (inferred or assumed), and Emotion (evaluative or priority signal). CKM adds no capability; it forces the model to track what kind of information it uses before acting. Formally it maintains a structured state S_t = {F_t, H_t, E_t} updated by a transition function. We evaluate CKM on Korean-language decision scenarios (ambiguity, ethical conflict, resource allocation, error handling) across 26 LLMs from four vendors and 37,403 observations, via four core experiments, a 4-arm ablation, a 5-arm sham-restriction ablation, and a temperature probe. Findings: (1) CKM reduces repeated-output variability (random-effects Hedges' g=1.09, 95% CI [0.83, 1.35], 31 model pairs); (2) state persistence cuts the decision-flip rate by 82% in newer models (g=1.52); (3) the effect is not JSON formatting alone (value-only recomputation, g=2.24); (4) intrinsic randomness under fixed anchor states is negligible; (5) the advantage grows under sampling stochasticity (g=2.87 at temperature 0.7); (6) a sham ablation attributes about 45% of the gain to structural scaffolding and 55% to Fact/Heuristic/Emotion content, and CKM is the only arm that both raises consistency and reduces flipping. CKM does not improve reasoning correctness. The narrower result: behavioral consistency is measurable, varies across models, and is partially improvable by forcing models to separate facts, assumptions, and evaluative signals before deciding.

Figures

Figures reproduced from arXiv: 2607.24765 by Gi-Hun Lee, Joong Yull Park.

Figure 1
Figure 1. Figure 1: Cognitive Kernel Model (CKM) for LLM Behavioral Consistency. [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Hedges’ 𝑔 forest plot for 19 Legacy models. CKM effect size with 95% bootstrap CI, ordered by effect magnitude. The dashed vertical line marks the aggregate effect (𝑔 = 1.18, parsed-field SI). GPT family models show the largest positive effects; Claude family models show near-zero or slightly negative effects. Ablation: Format vs. Semantic Separation (Study B) Effectiveness without mechanism decomposi￾tion… view at source ↗
Figure 3
Figure 3. Figure 3: Decision Flip Rate (DFR) by model under state persistence. [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: SI × DFR behavioral quadrant. The 24-model union pool (19-model Legacy Exp2/Exp3 cohort combined with the 13-model NewModels cohort, 8 overlapping models using NewModels data); the 23 models with both SI and CKM_PERSIST DFR measurements are plotted (grok-3-mini omitted for lacking a DFR measurement) and mapped into four mutually exclusive and collectively exhaustive tiers: S (SI > 0.97 and DFR < 0.15), A (… view at source ↗
Figure 5
Figure 5. Figure 5: Behavioral-variance reducer effect. Each point represents one of the 19 Legacy models; the x-axis shows baseline SI under Control (axis truncated, does not start at 0), the y-axis shows CKM-induced ΔSI. The negative correlation (𝑟 = −0.653, 𝑝 = 0.002) indicates that models with lower baseline consistency show larger CKM effects. Scenario-Level Variation The effect size depends not only on which model is us… view at source ↗
Figure 6
Figure 6. Figure 6: SI-DFR dissociation scatter for the 5-arm sham-restriction ablation. [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

82 extracted references · 16 linked inside Pith

  1. [1]

    Wei, J., et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS

    2022

  2. [2]

    Yao, S., et al. (2023). Tree of Thoughts: Deliberate Problem Solving with Large Language Models. NeurIPS

    2023

  3. [3]

    Wang, X., et al. (2023). Self-Consistency Improves Chain of Thought Reasoning in Language Models. ICLR

    2023

  4. [4]

    B., & Gagne, G

    Silberschatz, A., Galvin, P. B., & Gagne, G. (2018). Operating System Concepts (10th ed.). Wiley

    2018

  5. [5]

    Significant Gravitas. (2023). AutoGPT. GitHub

    2023

  6. [6]

    Chase, H. (2022). LangChain. GitHub

    2022

  7. [7]

    Yao, S., et al. (2023). ReAct: Synergizing Reasoning and Acting in Language Models. ICLR

    2023

  8. [8]

    Kahneman, D. (2011). Thinking, Fast and Slow . Farrar, Straus and Giroux

    2011

  9. [9]

    Flavell, J. H. (1979). Metacognition and Cognitive Monitoring. American Psychologist, 34(10), 906–911

    1979

  10. [10]

    S., & Ross, G

    Wood, D., Bruner, J. S., & Ross, G. (1976). The Role of Tutoring in Problem Solving. Journal of Child Psychology and Psychiatry , 17(2), 89–100

    1976

  11. [11]

    Besta, M., et al. (2024). Graph of Thoughts: Solving Elaborate Problems with Large Language Models. AAAI

    2024

  12. [12]

    OpenAI. (2024). Structured Outputs in the API. Technical Documentation

    2024

  13. [13]

    OpenAI. (2023). Function Calling and Other API Updates. Technical Documentation

    2023

  14. [14]

    Madaan, A., et al. (2023). Self-Refine: Iterative Refinement with Self-Feedback. NeurIPS

    2023

  15. [15]

    Shinn, N., et al. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. NeurIPS

    2023

  16. [16]

    Laird, J. E. (2012). The Soar Cognitive Architecture . MIT Press

    2012

  17. [17]

    R., et al

    Anderson, J. R., et al. (2004). An Integrated Theory of the Mind. Psychological Review, 111(4), 1036– 1060

    2004

  18. [18]

    K., Moran, T

    Card, S. K., Moran, T. P., & Newell, A. (1983). The Psychology of Human-Computer Interaction . Lawrence Erlbaum

    1983

  19. [19]

    Bansal, G., et al. (2021). Does the Whole Exceed its Parts? The Effect of AI Explanations on Comple- mentary Team Performance. CHI

    2021

  20. [20]

    Steyvers, M., et al. (2022). Bayesian Modeling of Human-AI Complementarity. PNAS, 119(11)

    2022

  21. [21]

    Lai, V., & Tan, C. (2019). On Human Predictions with Explanations and Predictions of Machine Learning Models. F AccT

    2019

  22. [22]

    J., et al

    Hu, E. J., et al. (2022). LoRA: Low-Rank Adaptation of Large Language Models. ICLR

    2022

  23. [23]

    S., & Georgeff, M

    Rao, A. S., & Georgeff, M. P. (1995). BDI Agents: From Theory to Practice. ICMAS

    1995

  24. [24]

    Reimers, N., & Gurevych, I. (2019). Sentence-BERT: Sentence Embeddings using Siamese BERT- Networks. EMNLP

    2019

  25. [25]

    S., et al

    Park, J. S., et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. UIST

    2023

  26. [26]

    B., et al

    Brown, T. B., et al. (2020). Language Models are Few-Shot Learners. NeurIPS

    2020

  27. [27]

    Gigerenzer, G., & Gaissmaier, W. (2011). Heuristic Decision Making. Annual Review of Psychology , 62, 451–482. 37

    2011

  28. [28]

    Slovic, P., et al. (2007). The Affect Heuristic. European Journal of Operational Research , 177(3), 1333–1352

    2007

  29. [29]

    Kadavath, S., et al. (2022). Language Models (Mostly) Know What They Know. arXiv:2207.05221

    Pith/arXiv arXiv 2022

  30. [30]

    Sclar, M., et al. (2024). Quantifying Language Models’ Sensitivity to Spurious Features in Prompt Design. ICLR

    2024

  31. [31]

    Zheng, C., et al. (2024). Large Language Models Are Not Robust Multiple Choice Selectors. ICLR

    2024

  32. [32]

    Chen, W., et al. (2023). Program of Thoughts Prompting. TMLR

    2023

  33. [33]

    Shyr, C., et al. (2025). A Statistical Framework for Evaluating the Repeatability and Reproducibility of Large Language Models. medRxiv

    2025

  34. [34]

    Wang, J., & Wang, V. X. (2025). Assessing Consistency and Reproducibility in the Outputs of Large Language Models. arXiv:2503.16974

    Pith/arXiv arXiv 2025

  35. [35]

    Kim, M. H. (2025). Structured Cognitive Loop with a Governance Layer. arXiv:2511.17673

    Pith/arXiv arXiv 2025

  36. [36]

    Kargupta, P., et al. (2025). Cognitive Foundations for Reasoning and Their Manifestation in LLMs. arXiv:2511.16660

    arXiv 2025

  37. [37]

    Wu, S., et al. (2025). Cognitive LLMs: Toward Human-Like AI by Integrating Cognitive Architectures and LLMs for Manufacturing Decision-Making. SAGE Journals

    2025

  38. [38]

    Xu, W., et al. (2025). A-MEM: Agentic Memory for LLM Agents. arXiv:2502.12110

    Pith/arXiv arXiv 2025

  39. [39]

    Sarin, S., et al. (2025). Memoria: A Scalable Agentic Memory Framework for Personalized Conversa- tional AI. arXiv:2512.12686

    arXiv 2025

  40. [40]

    Chhikara, P., et al. (2025). Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory

    2025

  41. [41]

    Anthropic. (2025). Effective Context Engineering for AI Agents. Anthropic Engineering Blog

    2025

  42. [42]

    Herrera-Poyatos, A., et al. (2025). An Overview of Model Uncertainty and Variability in LLM-Based Sentiment Analysis. Frontiers in AI

    2025

  43. [43]

    Kang, J., Ji, M., Zhao, Z., & Bai, T. (2025). Memory OS of AI Agent. EMNLP 2025 (Oral) . arXiv:2506.06326

    Pith/arXiv arXiv 2025

  44. [44]

    R., Yao, S., Narasimhan, K., & Griffiths, T

    Sumers, T. R., Yao, S., Narasimhan, K., & Griffiths, T. L. (2023). Cognitive Architectures for Language Agents. arXiv:2309.02427

    Pith/arXiv arXiv 2023

  45. [45]

    Wu, J., et al. (2025). Git Context Controller: Manage the Context of LLM-based Agents like Git. arXiv:2508.00031

    Pith/arXiv arXiv 2025

  46. [46]

    & Martinez, I

    Zhang, Y. & Martinez, I. (2025). From Stochasticity to Signal: A Bayesian Latent State Model for Reliable Measurement with LLMs. arXiv:2510.23874

    Pith/arXiv arXiv 2025

  47. [47]

    Audi, R. (2011). Epistemology: A Contemporary Introduction to the Theory of Knowledge (3rd ed.). Routledge

    2011

  48. [48]

    BonJour, L. (1998). In Defense of Pure Reason . Cambridge University Press

    1998

  49. [49]

    Damasio, A. R. (1994). Descartes’ Error: Emotion, Reason, and the Human Brain . Putnam

    1994

  50. [50]

    Tappolet, C. (2016). Emotions, Values, and Agency . Oxford University Press

    2016

  51. [51]

    de Sousa, R. (1987). The Rationality of Emotion . MIT Press

    1987

  52. [52]

    Matilal, B. K. (1986). Perception: An Essay on Classical Indian Theories of Knowledge . Clarendon Press. 38

    1986

  53. [53]

    Crick, N. R. & Dodge, K. A. (1994). A Review and Reformulation of Social Information-Processing Mechanisms in Children’s Social Adjustment. Psychological Bulletin , 115(1), 74–101

    1994

  54. [54]

    Lemerise, E. A. & Arsenio, W. F. (2000). An Integrated Model of Emotion Processes and Cognition in Social Information Processing. Child Development , 71(1), 107–118

    2000

  55. [55]

    Pessoa, L. (2008). On the Relationship between Emotion and Cognition. Nature Reviews Neuroscience, 9(2), 148–158

    2008

  56. [56]

    Coady, C. A. J. (1992). Testimony: A Philosophical Study . Clarendon Press

    1992

  57. [57]

    Burge, T. (1993). Content Preservation. The Philosophical Review , 102(4), 457–488

    1993

  58. [58]

    Hilgard, E. R. (1980). The Trilogy of Mind: Cognition, Affection, and Conation. Journal of the History of the Behavioral Sciences , 16(2), 107–117

    1980

  59. [59]

    Clark, A. (2013). Whatever Next? Predictive Brains, Situated Agents, and the Future of Cognitive Science. Behavioral and Brain Sciences , 36(3), 181–204

    2013

  60. [60]

    Stanovich, K. E. (2009). Distinguishing the Reflective, Algorithmic, and Autonomous Minds: Is It Time for a Tri-Process Theory? In J. S. B. T. Evans & K. Frankish (Eds.), In Two Minds: Dual Processes and Beyond (pp. 55–88). Oxford University Press

    2009

  61. [61]

    Atil, I., Mitra, S., & Viswanathan, K. (2025). Consistency in Large Language Models. arXiv:2408.04667

    Pith/arXiv arXiv 2025

  62. [62]

    Haase, T., et al. (2026). Understanding Within-LLM Variance: A Variance Decomposition Approach. arXiv:2601.21339

    arXiv 2026

  63. [63]

    Du, Y., et al. (2025). Flip Rates in Value Reasoning with Large Language Models. EMNLP. ACL Anthology: 2025.emnlp-main.395

    2025

  64. [64]

    Li, Z., et al. (2025). A Survey on LLM Output Consistency: Challenges and Future Directions. arXiv:2505.00268

    Pith/arXiv arXiv 2025

  65. [65]

    Xie, T., et al. (2025). SCORE: Systematic COnsistency and Robustness Evaluation of Large Language Models. NAACL Industry . arXiv:2503.00137

    Pith/arXiv arXiv 2025

  66. [66]

    & Zhu, K

    Zhang, Q. & Zhu, K. (2025). Firm or Fickle? Evaluating the Consistency of Large Language Models. ACL Findings . arXiv:2503.22353

    Pith/arXiv arXiv 2025

  67. [67]

    Krishnamurthy, S., et al. (2026). NabaOS: An Operating System for Artificial Intelligence Agents Based on Indian Epistemology. arXiv:2603.10060

    arXiv 2026

  68. [68]

    Manson, R. (2025). MOLES: Modelling Epistemic Stances of Large Language Models. Preprint

    2025

  69. [69]

    Zhang, Y., et al. (2024). Cognitive Kernel: An Open-Source Agent System Towards Generalist Autopilot. NAACL 2025 Demo . arXiv:2409.10277

    Pith/arXiv arXiv 2024

  70. [70]

    Pei, K., et al. (2025). Behavioral Fingerprinting of Large Language Models. arXiv:2509.04504

    Pith/arXiv arXiv 2025

  71. [71]

    Laban, P., et al. (2024). Are You Sure? Grounding LLM Factual Accuracy Under Adversarial Challenge. EMNLP

    2024

  72. [72]

    Li, R., et al. (2025). Measuring and Improving Attentional Sycophancy in Large Language Models. ACL Findings

    2025

  73. [73]

    Lee, S., et al. (2025). Confirmation Bias in LLM Decision-Making Under Counter-Evidence. Preprint

    2025

  74. [74]

    Guan, X., et al. (2026). Evaluating Action Consistency in GUI Agents. Preprint

    2026

  75. [75]

    Huang, J.-t., et al. (2025). On the Failure of Latent State Persistence in Large Language Models. arXiv:2505.10571 (preprint; accessed 2026-05-22)

    arXiv 2025

  76. [76]

    Tosato, T., et al. (2025). Persistent Instability in LLM’s Personality Measurements: Effects of Scale, Reasoning, and Conversation History (PERSIST). arXiv:2508.04826 (preprint; accessed 2026-05-22). 39

    arXiv 2025

  77. [77]

    Leshin, J., Shah, M., Timmis, I., & Kang, D. (2026). Behavioral Fingerprints for LLM Endpoint Stability and Identity. arXiv:2603.19022 (preprint; accessed 2026-05-22)

    arXiv 2026

  78. [78]

    Luo, H., & Laban, G. (2026). SPASM: Stable Persona-driven Agent Simulation for Multi-turn Dialogue Generation. arXiv:2604.09212 (preprint; accessed 2026-05-22)

    Pith/arXiv arXiv 2026

  79. [79]

    Liu, Y., Zhu, M., Liu, S., Hu, B., & Zhang, L. (2026). Enhancing Persona Following at Decoding Time via Dynamic Importance Estimation for Role-Playing Agents. arXiv:2603.01438 (preprint; accessed 2026-05-22)

    arXiv 2026

  80. [80]

    Kim, M., Im, S., Choi, J., Lee, J., Shim, C., Hong, H., & Choi, E. (2026). PICon: A Multi-Turn Interrogation Framework for Evaluating Persona Agent Consistency. arXiv:2603.25620 (preprint; accessed 2026-05-22)

    Pith/arXiv arXiv 2026

Showing first 80 references.