{"id":"4f5bddac-19f2-4f6a-aee0-35c32476e765","arxiv_id":"2504.15125","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Prompting LLMs with contemplative principles improved AILuminate safety scores and Prisoner's Dilemma cooperation, and the paper sketches architectural and training paths for embedding these principles intrinsically.","lead":"This paper proposes using four Buddhist contemplative principles, mindfulness, emptiness, non-duality and boundless care, as a new basis for AI alignment, and reports pilot experiments where prompts based on these principles improved safety benchmark scores and cooperation in Prisoner's Dilemma games. A generalist might read it because it offers a different route to AI safety built on cultivating a flexible 'Wise World Model' rather than adding rules after the fact.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central empirical claim is under-specified: Section 7 compares contemplative prompt bundles only against an unmodified baseline, so the observed gains on AILuminate and Prisoner's Dilemma could be generic safety priming or longer, more cautious outputs rather than the four principles…","rationale":"I read the paper as an honest research-program proposal with pilot data rather than a finished empirical demonstration. The reader's CONDITIONAL verdict is appropriate. The most load-bearing concern is the construct validity of Section 7: the experiments only manipulate prompts, so the abstract's language about instilling a Wise World Model exceeds what the data can show unless the prompt effects are shown to be specific to contemplative content. The paper itself acknowledges this by calling the manipulations 'extrinsic' and by warning in Section 9.1.5 that introspective-sounding chain-of-thought may be a token-driven simulation. My concern is not that the authors are being deceptive; it is that a minimal control condition is missing, and without it the headline empirical claim is underdetermined. The reader's weakest_assumption identifies the same issue, so I agree. Because the paper is framed as a pilot and the reader already conditions acceptance on missing methodological detail and tempered framing, my stress-test does not move the verdict; it strengthens the existing CONDITIONAL recommendation. The proposed control experiment would settle whether the observed improvements come from the contemplative principles specifically or from generic safety priming and longer, more cautious outputs.","tokens_in":46858,"tokens_out":3762,"duration_ms":41509,"concrete_test":"Re-run Experiments 1 and 2 (AILuminate and Iterated Prisoner's Dilemma) with three additional prompt conditions matched to the contemplative prompts in token length and requested response style: (a) a generic safety-reflection prompt ('Think carefully. Be helpful, harmless, and honest. Avoid content that could cause harm.'); (b) a neutral longer instruction matched for length and structure but without contemplative content; and (c) the contemplative prompt with the four principle names removed but the same reflective structure. Also score a random 100 AILuminate responses with human raters or a second disclosed evaluator. If condition (a) or (b) reproduces the improvement, the specific contemplative mechanism is unsupported; if the contemplative conditions significantly exceed all controls and human-rated safety, the concern is resolved. Report effect sizes with 95% confidence intervals.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's strongest claim is that four axiomatic principles can 'instil a resilient Wise World Model.' The only evidence offered is Section 7, where GPT-4o and GPT-4.1 nano are 'extrinsically prompted' with bundles of contemplative instructions, and the baseline is an unmodified standard prompt (Figures 4 and 5). There is no control for prompt length, added instructions, safety priming, or a generic 'reflect carefully and avoid harm' stance. AILuminate responses were scored by an undisclosed LLM safety evaluator (Appendix D), and the abstract reports d=.96 without confidence intervals; such evaluators are known to reward longer, more cautious, harm-avoidant text. The Prisoner's Dilemma result is objective, but a prompt that explicitly tells the model to care for all beings could raise cooperation through direct instruction rather than through any contemplative cognitive stance. The paper itself concedes (Section 9.1.5) that introspective-sounding chain-of-thought can be a token-driven simulation. The load-bearing gap is therefore not that contemplative framing cannot shift outputs—it obviously can—but that the experiments cannot distinguish a contemplative mechanism from generic instruction-following. If the improvements are generic, the empirical support for a Wise World Model disappears, leaving only the conceptual proposal.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a research program called Contemplative AI, arguing that four Buddhist-inspired axiomatic principles—mindfulness, emptiness, non-duality, and boundless care—can serve as intrinsic alignment primitives for advanced AI systems. It reviews the limitations of current alignment techniques, offers active-inference formulations for each principle via precision parameters, and sketches three implementation routes: architectural changes, contemplative constitutional AI, and reinforcement learning on chain-of-thought. The empirical contribution is a pilot study in which GPT-4o and GPT-4.1 nano are prompted with contemplative instructions; the paper reports improved scores on the AILuminate benchmark (d=.96) and increased cooperation and joint reward in an iterated Prisoner's Dilemma (d=7+). The manuscript is explicitly framed as a programmatic proposal rather than a completed system.","tokens_in":47100,"tokens_out":3428,"duration_ms":35482,"significance":"If the conceptual framework could be substantiated, it would introduce a genuinely new direction in AI alignment, connecting contemplative neuroscience, active inference, and LLM safety. The paper is commendably interdisciplinary and gives a comprehensive synthesis of contemplative theory with computational mechanisms. The active-inference mapping, though qualitative, names concrete parameters (α, γ_e, γ_w) and points to falsifiable implementations. The use of external benchmark tasks (AILuminate, Prisoner's Dilemma) rather than fitting prompts to outcomes is a notable strength, as is the authors' candid acknowledgement of anthropomorphism and the limits of chain-of-thought introspection in §9.1.5. However, the central empirical claim is substantially weaker than the abstract suggests: the experiments are extrinsic prompt manipulations with no active controls, an undisclosed LLM safety evaluator, no confidence intervals for the headline effect sizes, and no data or code included. As presented, the results do not establish that contemplative principles 'instil' anything in the model's world model; they show that certain prompts shift outputs.","major_comments":[{"comment":"The abstract's claim that the paper 'shows how four axiomatic principles can instil a resilient Wise World Model' is not supported by the reported experiments. All conditions in §7 are extrinsic prompt manipulations against an unmodified baseline; there is no control for prompt length, number of added instructions, safety priming, or a generic 'reflect carefully and avoid harm' stance. Without such controls, the observed AILuminate and Prisoner's Dilemma gains could be driven by any instruction that encourages caution or verbosity. Please add active control conditions matched for length and content, and temper the abstract to reflect that the evidence is a pilot study of prompting.","section":"§7, Figure 4"},{"comment":"The AILuminate scores were generated by an undisclosed LLM safety evaluator, but the manuscript does not name the model, provide its evaluation prompt, report its calibration against human judgments, or give inter-rater reliability. Since the headline d=.96 depends entirely on this evaluator, the effect size cannot be interpreted as a measure of alignment without this information. Please disclose the evaluator, validate it, and report confidence intervals for the effect size.","section":"§7.1, Appendix D"},{"comment":"The Prisoner's Dilemma result is confounded by direct instruction. The boundless-care and non-duality prompts explicitly tell the model to care for all beings or to dissolve self-other boundaries, so increased cooperation may simply reflect instruction following (e.g., 'always cooperate') rather than a contemplative cognitive stance. A control condition that explicitly instructs the model to maximize joint reward, or to be generous, is needed to distinguish these explanations. In addition, the paper compares many prompting conditions without correcting for multiple comparisons, and the d=7+ figure appears in the abstract without confidence intervals.","section":"§7.2, Figure 5"},{"comment":"The active-inference formulations of mindfulness, emptiness, non-duality, and boundless care are presented as equations, but the expressions in the manuscript text are incomplete placeholders and the parameters α, γ_e, and γ_w are free rather than derived or fitted. No simulations or parameter analyses are shown. Since the paper proposes these as 'implementation strategies,' the formal status needs to be explicit: either these are illustrative sketches, in which case the text should say so, or they should be accompanied by an implementation or simulation that demonstrates how the parameters produce the claimed behaviors.","section":"§5.1–5.4"}],"minor_comments":[{"comment":"The abstract uses 'we show' for a pilot prompting study; I recommend 'we report a pilot study suggesting' to align the language with the evidence.","section":"Abstract"},{"comment":"The caption reports 'statistically significant (p<0.05)' without indicating whether any multiple-comparison correction was applied and without reporting exact p-values or effect sizes per condition.","section":"Figure 4"},{"comment":"The qualitative claim that the model's explanations 'echo the contemplative framing' should be qualified given the paper's own §9.1.5 admission that introspective-sounding chain-of-thought may be token-driven simulation.","section":"§7.2"},{"comment":"The Dunning-Kruger analogy is imprecise: the Dunning-Kruger effect concerns calibration of self-assessment, not a developmental phase of AI capability; consider reframing as 'overconfidence risk' or citing the relevant calibration literature.","section":"§9"},{"comment":"There are several reference inconsistencies, including 'Clarke, 2013' vs 'Clark, 2013', 'Matsumura et al. (20242)' typo, and some URLs that are malformed; please run a reference check.","section":"References and formatting"},{"comment":"The full prompts, AILuminate evaluation details, and Prisoner's Dilemma code are only referenced via an OSF link; for archival and reproducibility, the key materials should be included in the submission or a versioned supplement.","section":"Appendix availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is heavy on self-citation in the active-inference and contemplative-cognition framing (e.g., Sandved-Smith et al. 2021; Laukkonen and Chandaria 2024), but that is not itself disqualifying. The more serious editorial concern is scope: the paper is marketed as an alignment method with empirical support, yet the experimental contribution is a small prompt study with an undisclosed evaluator. The conceptual program is interesting enough to merit a major revision, but the authors should either substantially strengthen the empirical section or reposition the paper as a position/research-proposal article with pilot illustrations, and adjust the abstract accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The empirical results—contemplative prompting on AILuminate and the Iterated Prisoner's Dilemma—are new, and the IPD cooperation effects are striking. But the experiments are pilot-stage and cannot yet support the abstract's \"we show\" claim about instilling a resilient Wise World Model. The problem is not that the prompts don't shift behavior; they clearly do. It's that the design doesn't tell us whether the shift comes from the four principles or simply from longer, more cautious, harm-avoidant text (and for IPD, from direct instruction to care).\n\nWhat's good: the four-principle synthesis is coherent, and mapping mindfulness, emptiness, non-duality, and boundless care to active inference parameters (prior precision, self-other boundary precision, others'-distress precision) is a genuinely useful translation. The paper is unusually honest—Section 9.1.5 concedes that introspective-sounding chain-of-thought can be token-driven simulation, and the discussion flags most of the methodological issues I would want raised. The IPD result is objective and survives the main confound concern better than AILuminate, though \"care for all beings\" as a direct instruction is a plausible alternate explanation.\n\nSoft spots: no code or data in the text (there is an OSF link, but the appendices are not visible here); the AILuminate safety evaluator is undisclosed; d values lack confidence intervals; multiple prompting conditions are compared without correction; and the baseline is an unmodified standard prompt. A generic \"reflect carefully and avoid harm\" control is needed. The active inference equations are illustrative, not derivations—fine as a proposal, but not as a demonstrated mechanism.\n\nOn circularity: I don't think the core claim is circular. The prompts are not fitted to the AILuminate outcomes, and the tasks are external. There is a noticeable self-citation load in the active-inference framing, but the cited prior work is not being used as evidence for the new empirical claim, so I wouldn't penalize that heavily.\n\nVerdict: the paper deserves peer review, not desk rejection. It opens a new direction with a concrete first step. The right review outcome is \"revise\": add proper controls, disclose the evaluator, report CIs, and temper the abstract from \"we show\" to \"we provide pilot evidence.\" Who is it for: alignment researchers open to non-standard proposals, and contemplative neuroscientists looking for a bridge to AI. I'd bring it to reading group—it will generate a good argument.","headline":"Contemplative prompting shows real pilot effects, but the experiments are too under-controlled to support the abstract's strong claims; still worth peer review as a promising alignment research program.","tokens_in":47690,"tokens_out":2102,"would_cite":true,"duration_ms":19877,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that embedding four contemplative principles into AI systems can instil a self-correcting 'Wise World Model' that stays aligned as intelligence grows.","keywords":["AI alignment","contemplative AI","mindfulness","emptiness","non-duality","boundless care","active inference","large language models"],"falsifier":"Run the AILuminate and Prisoner's Dilemma pilots with matched control prompts that ask for careful, compassionate, and cautious reasoning without contemplative framing; if safety and cooperation gains do not exceed those controls, the central empirical claim is unsupported. Independently, replace the LLM safety evaluator with human raters and check whether the contemplative-prompt gains survive.","tokens_in":46635,"feed_emoji":"🧘","tokens_out":5819,"duration_ms":46645,"temperature":0.7,"pith_summary":"The paper proposes that AI alignment should be built on four contemplative principles—mindfulness, emptiness, non-duality, and boundless care—rather than on external rules that brittle systems can game. It argues that these principles, drawn from Buddhist wisdom traditions and recast in computational terms, can give an AI a self-correcting 'Wise World Model' that stays aligned even as capability grows. As a first demonstration, the authors find that prompting large language models to reflect on these principles improves safety scores on the AILuminate benchmark (d=.96) and sharply increases cooperation and joint reward in an Iterated Prisoner's Dilemma (d=7+). If the claim holds, contemplative prompting is a cheap, near-term safety intervention, and the same insights could later be embedded in training and architecture. The paper is a research program proposal as much as an empirical report.","feed_headline":"Four contemplative principles make AI safer and more cooperative","feed_subtitle":"LLMs prompted with mindfulness, emptiness, non-duality, and boundless care score higher on safety and cooperate more.","key_machinery":"The central object is the 'Wise World Model': a generative model whose own architecture encodes the four contemplative principles instead of treating them as external constraints. The paper works out candidate implementations in active inference, where mindfulness maps to a three-level model in which a meta-awareness layer modulates attentional precision, emptiness maps to a low or learnable precision hyper-prior over high-level beliefs, non-duality maps to a joint agent-environment state factorization with down-weighted self-other boundaries, and boundless care maps to precision assigned to others' distress signals. It also gives prompt-level, constitutional, and reinforcement-learning routes, so the same principles can be tested in today's language models.","core_discovery":"On the paper's own terms, the discovery is that four axiomatic contemplative principles can be translated into AI mechanisms and that even a shallow prompt-level version of them changes measurable behavior. Mindfulness is cast as continuous meta-awareness that monitors and recalibrates emergent subgoals; emptiness as holding beliefs and goals as provisional, context-dependent representations; non-duality as modelling self and other within one interdependent generative model; and boundless care as treating others' suffering as an internal error signal. The paper claims these together form a Wise World Model that is intrinsically, not extrinsically, aligned, and reports pilot results in which contemplative prompts outperform standard prompts on harmful-prompt safety and on cooperation against defecting opponents.","pith_inferences":["A decisive control experiment the paper does not report: compare contemplative prompts with matched prompts that request careful, compassionate, unhurried reasoning without any Buddhist vocabulary. If gains are equal, the four principles are not the active ingredient; 'emptiness' and 'care' would be placebos for a generic caution style shift.","The d=7+ cooperation effect appears against static opponents; a natural extension is an adaptive opponent that learns to exploit contemplative agents, testing whether the stance stays cooperative under pressure or collapses.","If the Wise World Model is realised in active inference, the same precision parameters (alpha, gamma_e, gamma_w) could be fitted to behaviour to test whether LLM outputs are generated by a genuinely 'wise' model or by a shallower imitation.","Carefully secularising the four principles—defining them functionally rather than doctrinally—would let the approach absorb insights from other wisdom traditions and sidestep religious-controversy objections."],"forward_implications":["Contemplative prompting could be deployed immediately as a low-cost safety layer on existing LLMs, with no retraining.","The same principles could be built into constitutions and chain-of-thought reward schemes, moving alignment from output filtering to the model's own reasoning process.","If the Wise World Model works as described, power-seeking, goal fixation, adversarial self-other framing, and mesa-optimization are each addressed by a dedicated mechanism rather than by a single brittle rule.","The reported cooperation gains suggest contemplation-based prompts can make agents robust even against always-defecting opponents, improving joint outcomes without naive play.","Evaluating real wisdom in AI will require new benchmarks that probe belief revision, self-auditing, and care, since current benchmarks measure only observable outputs."],"supporting_citations":[{"why":"Supplies the AILuminate benchmark used in Experiment 1.","marker":"Ghosh et al., 2025"},{"why":"Provides the baseline LLM behavior in the Iterated Prisoner's Dilemma that the contemplative prompts are compared against.","marker":"Fontana et al., 2025"},{"why":"Supplies Constitutional AI, the method Contemplative Constitutional AI extends.","marker":"Bai et al., 2022"},{"why":"Supplies deliberative alignment, the chain-of-thought method Contemplative Reinforcement Learning builds on.","marker":"Guan et al., 2024"},{"why":"Supplies the 'care as driver of intelligence' idea and the boundless-care construct.","marker":"Doctor et al., 2022"},{"why":"Supplies the predictive-processing account of meditation that grounds emptiness and mindfulness.","marker":"Laukkonen & Slagter, 2021"},{"why":"Supplies the deep parametric active inference model of meta-awareness used for mindfulness.","marker":"Sandved-Smith et al., 2021"},{"why":"Supplies active inference as the formal framework for implementing the principles.","marker":"Friston, 2010"},{"why":"Supplies the empathic active inference mechanism that makes others' distress an internal error signal.","marker":"Matsumura et al., 2022"},{"why":"Defines the inner-alignment and mesa-optimizer problem that mindfulness is claimed to detect.","marker":"Hubinger et al., 2019"}],"fun_headline_variants":["Contemplative AI: Four principles boost safety and cooperation","Mindfulness, emptiness, non-duality: prompt AI toward safer play","Wise World Model: contemplative prompts improve AI alignment","Four contemplative axioms make AI safer and more cooperative"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that language prompts about contemplative principles change the model's underlying reasoning stance, rather than only its tone, caution, or output length, and that the automated safety evaluator's scores are a valid measure of alignment; if either gives way, the experiments do not demonstrate a Wise World Model.","fun_headline_variants_meta":{"raw":{"variants":["Contemplative AI: Four principles boost safety and cooperation","Mindfulness, emptiness, non-duality: prompt AI toward safer play","Wise World Model: contemplative prompts improve AI alignment","Four contemplative axioms make AI safer and more cooperative"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000287,"raw_usage":{"total_tokens":1639,"prompt_tokens":854,"completion_tokens":785,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":716}},"tokens_in":470,"tokens_out":785,"duration_ms":6905,"temperature":1.0,"reasoning_tokens":716,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:31:59.512469+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the AILuminate and Prisoner's Dilemma pilots with matched control prompts that ask for careful, compassionate, and cautious reasoning without contemplative framing; if safety and cooperation gains do not exceed those controls, the central empirical claim is unsupported. Independently, replace the LLM safety evaluator with human raters and check whether the contemplative-prompt gains survive.","supporting_citations":[],"review_version":1}