{"id":"a5e8e58d-462d-4dd2-8afe-9803c3e0badd","arxiv_id":"2607.10539","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A unified LLM pipeline with Bayesian trait updates, conformal sets, and periodic memory-anchor refresh improves calibration and long-horizon persona fidelity over static prompting on module benchmarks.","lead":"AI YOU continually infers a 22-dimension personality profile from chat using Bayesian updates and conformal prediction, then keeps a digital twin consistent via a refreshed three-layer memory. It is a practical recipe for less-drifty persona agents with reported uncertainty, aimed at companion and consulting twins.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Persona-fidelity claim rests on fictional/adversarial probes, not real-individual twins; Bayesian noise model is secondary.","rationale":"The reader correctly flags the absence of end-to-end longitudinal user studies and the fragility of treating single-pass LLM confidences as observation noise (§3.1). That noise model is a genuine soft spot for sequential Bayesian updates, but it is not the single most load-bearing concern for the paper’s central claim. The claim that is doing the work for “AI YOU Town / personal digital twin” is the persona-fidelity / trait-drift result in Table 7 and the Abstract’s item (iii). That result is measured only on fictional roles and multi-agent games with assigned targets; the paper never shows that the same pipeline improves fidelity to a real individual’s longitudinal behavior. Module-level calibration and memory ablations remain solid systems evidence and justify a CONDITIONAL verdict if claims stay scoped. I therefore keep CONDITIONAL (same as the reader) but re-center the weakest link on the fictional-to-real transfer rather than primarily on the σ²_obs formula. Agreement is partial: same overall risk level and verdict, different primary soft spot.","tokens_in":27456,"tokens_out":692,"duration_ms":29993,"concrete_test":"Hold out a small consented panel of real users with independent Big-Five (or 22-d) retest scores; run 100-turn twin dialogues under Refresh vs Static; compute trait MAD against the user’s retest profile and human-rated fidelity. If Refresh does not reduce MAD / raise fidelity relative to Static by a margin comparable to Table 7, the persona-fidelity half of the strongest claim does not transfer.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The Abstract and §4.7 claim that AI YOU “enhances persona fidelity over static prompting in role playing over 100 turns while reducing trait drift, for most evaluated backbones under adversarial settings.” That claim is load-bearing for the digital-twin framing, yet Table 7 measures only (a) LLM-as-judge scores on eight fictional Persistent Personas roles and (b) Big-Five MAD under a 7-agent Werewolf game with assigned trait targets. Neither setting is an individualized PDT: there is no longitudinal ground-truth profile of a real person, no user correction loop, and no test that the inferred 22-d state actually improves fidelity to that person. The Limitations section itself states that these benchmarks “do not by themselves validate fidelity for real individuals.” Module gains on PANDORA/Essays (small MAE, better ECE/coverage) and LoCoMo ablations therefore support calibrated state estimation and memory-grounded QA, but do not license the transfer to “personal digital twin” fidelity that the title and strongest claim assert. The reader’s weakest assumption (LLM confidences as σ²_obs) is real but secondary; even a perfect noise model would not close the fictional-to-real gap.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes AI YOU, a multi-module framework that infers a 22-dimensional user profile from dialogue via structured prompting, Gaussian conjugate Bayesian updates, and conformal prediction sets, then conditions persona-aware generation on a periodically refreshed memory anchor and a three-layer (working/episodic/semantic) memory. Auxiliary monitors track affect, relationship state, and scam/manipulation risk. Module-level experiments on PANDORA, Essays, DailyDialog, PsyScam, LoCoMo, and PersonaConflicts report modest MAE gains, consistent ECE reductions, conformal coverage 0.921–0.976, and ablation drops when Bayesian, conformal, memory, or risk components are removed. A separate persona-preservation study (Table 7) compares periodic refresh vs. static prompting on eight fictional roles over 100 turns and on a 7-agent Werewolf game, reporting higher LLM-as-judge fidelity and lower Big-Five MAD for most backbones. A prototype AI YOU Town marketplace is described as an imaginative bidirectional twin environment.","tokens_in":27829,"tokens_out":836,"duration_ms":9996,"significance":"If the results hold, the work offers a practical, training-free recipe for calibrated sequential persona state estimation and long-horizon consistency that is more inspectable than static system prompts. Strengths include distribution-free conformal coverage above the nominal 90% target across multiple backbones (Table 2), ablations that generally move metrics in the expected direction (Bayesian, conformal, memory, risk, context), and an explicit Limitations section that flags the absence of end-to-end longitudinal user studies. The combination of Bayesian evidence weighting with Adaptive Prediction Sets for LLM trait estimates is a useful engineering contribution for digital-twin research, even if the transfer to real-individual fidelity remains unproven.","major_comments":[{"comment":"Abstract and §4.7 claim that AI YOU “enhances persona fidelity … while reducing trait drift … under adversarial settings,” which is load-bearing for the personal-digital-twin framing. Table 7 only evaluates (a) LLM-as-judge scores on eight fictional Persistent Personas roles and (b) Big-Five MAD under assigned trait targets in a 7-agent Werewolf game. Neither setting has a longitudinal ground-truth profile of a real individual, a user correction loop, or a test that the inferred 22-d state improves fidelity to that person. The Limitations section itself states that these benchmarks “do not by themselves validate fidelity for real individuals.” The claim should be narrowed to “fictional/adversarial role consistency” or supported by a real-user longitudinal probe before the twin framing is retained at full strength.","section":null},{"comment":"§3.1 defines observation variance as σ²_obs = max(10^{-3}, 1−c_t) and feeds single-pass LLM confidences into conjugate Gaussian updates. Table 2 shows ECE drops and coverage above 90%, but there is no calibration diagnostic of whether c_t is a valid noise scale (e.g., reliability diagrams of c_t vs. absolute error, or sensitivity of posterior MAE/coverage to alternative maps). Without that check, the sequential-update story rests on an unvalidated free parameter; either add the diagnostic or present Bayesian updating as a heuristic stabilizer rather than a calibrated likelihood model.","section":null},{"comment":"§4.1–4.6 and Appendix A.1 use fixed random subsamples (seed 42) and, for LoCoMo, a 300-instance diagnostic set whose absolute scores are not comparable to the full N=1542 run (Table 6 vs. Table 3). Several API rows also have JSON success <0.98 and are marked diagnostic. The paper should report confidence intervals or bootstrap variability for the main MAE/ECE/coverage numbers and clarify which claims rest only on diagnostic subsets, so that the “across main results” summary in the Abstract is not overstated.","section":null}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful core here is not a new twin theory. It is a working pipeline that treats persona as sequential state: LLM observations, Gaussian conjugate updates with confidence-as-noise, conformal sets (coverage 0.921–0.976), and a periodically refreshed three-layer memory anchor that actually reduces long-horizon drift on most backbones.\n\nWhat is new is the measured integration, not the parts. Bayesian preference updates, conformal NLP, Park-style working/episodic/semantic memory, and prompt personas are all cited prior art. The paper’s contribution is the 22-dim scaffold plus monitors (affect, risk, relationship) and the empirical package: multi-backbone tables, ablations that mostly move the right way, and 100-turn refresh vs static comparisons. Table 2 is the strongest evidence—ECE drops are consistent and larger than the modest MAE gains; conformal coverage sits above the 90% target. Memory ablations on LoCoMo show retrieval is load-bearing. Risk and context ablations are backbone-dependent in a way the authors report honestly. Limitations and ethics are unusually clear for a systems demo with a marketplace hook.\n\nThe soft spot is claim scope, not math fraud. Abstract and §4.7 sell “personal digital twin” fidelity and reduced trait drift under adversarial multi-agent settings. Table 7 is LLM-as-judge on eight fictional roles plus Big-Five MAD in a Werewolf game with assigned targets. That is a fair stress test of prompt drift; it is not fidelity to a real individual’s longitudinal profile. The Limitations section already says so. The σ²_obs = max(10^{-3}, 1−c_t) map is a free design choice and secondary; even a perfect noise model would not close the fictional-to-real gap. Title/marketplace framing (“Make Friends and Money”) outruns the measured science. Free parameters (α, k, subsample seed) are normal for this genre but should be frozen and released with code.\n\nWho it is for: people building companion agents, long-horizon personalization, or multi-agent persona sims who need calibrated state rather than another static system prompt. Citation pattern is appropriate; circularity burden is low because external corpora and independent probes are used.\n\nI would send it to peer review as a systems contribution if claims stay scoped to module calibration, memory-grounded reasoning, and simulated persona stability. I would not treat it as validated real-individual twinning until longitudinal user studies and code exist. Engage the methods and tables; discount the twin-town rhetoric.","headline":"Solid integrated systems paper on calibrated persona state + memory-anchor refresh; the digital-twin fidelity claim is oversold relative to fictional/adversarial probes.","tokens_in":28423,"tokens_out":606,"would_cite":true,"duration_ms":7782,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A digital twin can keep a 22-dimensional personality profile updated from conversation, with calibrated uncertainty and long-horizon consistency, instead of static persona prompts that drift.","keywords":["personal digital twin","persona inference","Bayesian updating","conformal prediction","three-layer memory","persona consistency","uncertainty calibration","multi-agent role-play"],"falsifier":"Run a consented multi-week user study in which the same people both self-report traits and talk to AI YOU twins: if conformal coverage falls below the nominal 90% target, or if trait drift and judge-scored persona fidelity under 100-turn adversarial probing are no better than a static persona prompt with full history, the central claim fails.","tokens_in":28346,"feed_emoji":"🪞","tokens_out":950,"duration_ms":11272,"temperature":0.7,"pith_summary":"Static persona prompting gives large language models a fixed character card, but that card has no calibrated uncertainty, does not accumulate evidence turn by turn, and often drifts over long chats. This paper argues that a personal digital twin should instead be a living state: a 22-dimensional profile continually inferred from dialogue, updated with Bayesian belief revision, and wrapped in conformal prediction sets so the system knows what it does not know. That profile is embodied through a periodically refreshed memory anchor and a three-layer cognitive memory (working, episodic, semantic) that grounds generation without dumping the full history into the context window. Across Big Five prediction, affect and risk monitors, long-session memory QA, and 100-turn role-play under adversarial multi-agent pressure, the framework improves calibration and persona fidelity for most evaluated model backbones. The prototype AI YOU Town sketches a world in which such twins can socialize and work on a user’s behalf while remaining inspectable and uncertainty-aware.","feed_headline":"Digital twins that update personality without drifting","feed_subtitle":"Bayesian updates, conformal sets, and three-layer memory keep 100-turn personas calibrated.","key_machinery":"The AI YOU pipeline: structured prompting extracts candidate traits with confidences; Gaussian conjugate Bayesian updates (observation variance from 1−confidence) accumulate sequential evidence; conformal Adaptive Prediction Sets (target α=0.10) produce per-dimension prediction sets; a memory anchor refreshed every ~10 turns plus working/episodic/semantic layers stabilizes persona-conditioned generation.","core_discovery":"The authors claim that personal digital twins work when personality inference, sequential belief updating, uncertainty sets, and persona-conditioned generation are closed into one loop: conversation yields observations, Bayesian and conformal machinery maintain a calibrated 22-field profile, and a refreshed three-layer memory keeps the twin’s behavior aligned over 100-turn interactions better than static prompting.","pith_inferences":["If confidence-as-noise is mis-specified, the same pipeline could systematically under- or over-update traits for sparse or socially desirable text, so real deployments would need user correction and consent gates before any high-stakes use.","Closing the loop between twin simulation and profile update suggests multi-twin towns could become living preference laboratories—if identity and impersonation safeguards keep pace.","Because gains are stronger on API backbones than on smaller local models, practical twins may need backbone-specific monitor complexity rather than one universal scaffold."],"forward_implications":["Persona systems can report prediction sets and nulls instead of forced high-confidence trait labels when evidence is thin.","Periodic memory-anchor refresh can reduce style and knowledge drift over 100-turn role-play without per-persona fine-tuning.","Affect, relationship, and risk monitors can share the same three-layer memory so safety and personalization update together.","Digital-twin town prototypes can condition marketplace-style twin routing on calibrated, privacy-filtered state rather than a one-shot prompt.","Module ablations imply retrieval is the main driver of long-session memory QA; layer weighting remains backbone-dependent."],"fun_headline_variants":["Digital twins stay on persona with Bayesian updates and memory layers","Calibrated 22-dimension profiles keep twins consistent over 100 turns","Conformal sets and three-layer memory reduce trait drift in long chats","Closed loop of updates holds digital twins true over 100-turn roles","Refreshed memory anchors cut persona drift versus static prompting"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The system treats single-pass language-model confidence scores as usable observation noise for sequential Bayesian updates, and treats module-level benchmark gains plus simulated 100-turn role-play as evidence that the same loop will stay faithful for real individuals.","fun_headline_variants_meta":{"raw":{"variants":["Digital twins stay on persona with Bayesian updates and memory layers","Calibrated 22-dimension profiles keep twins consistent over 100 turns","Conformal sets and three-layer memory reduce trait drift in long chats","Closed loop of updates holds digital twins true over 100-turn roles","Refreshed memory anchors cut persona drift versus static prompting"]},"model":"grok-4.5","effort":"low","cost_usd":0.005858,"raw_usage":{"total_tokens":1479,"prompt_tokens":754,"num_sources_used":0,"completion_tokens":92,"cost_in_usd_ticks":58580000,"prompt_tokens_details":{"text_tokens":754,"audio_tokens":0,"image_tokens":0,"cached_tokens":0},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":633,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":754,"tokens_out":92,"duration_ms":5616,"temperature":1.0,"reasoning_tokens":633,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T10:57:49.573467+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run a consented multi-week user study in which the same people both self-report traits and talk to AI YOU twins: if conformal coverage falls below the nominal 90% target, or if trait drift and judge-scored persona fidelity under 100-turn adversarial probing are no better than a static persona prompt with full history, the central claim fails.","supporting_citations":[],"review_version":1}