{"id":"62b75462-6e4d-4d86-8f6e-394a27812135","arxiv_id":"2508.19942","paper_version":1,"verdict":"UNVERDICTED","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper proposes Socially Interactive Agents, powered by LLMs, RAG, and chain-of-thought prompting, as scalable facilitators for eliciting and transferring tacit knowledge in organizations, without providing empirical validation.","lead":"This paper proposes using artificial agents with conversational and social skills to draw out and preserve the hard-to-articulate, experience-based knowledge of employees before they leave an organization. It is a position paper: it describes the idea, the technical pieces, and the risks, but does not test whether the approach works.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Core concern: the paper's central mechanism assumes tacit knowledge can be externalized through natural-language dialogue, yet Section 2.1 defines tacit knowledge as 'difficult to formalize' and 'deeply embedded in actions'; no evidence supports this step.","rationale":"The reader's verdict of UNVERDICTED is appropriate because the paper is a conceptual proposal with no empirical or formal validation. My concern does not change that verdict: the central claim remains unverified, and my analysis identifies an additional conceptual gap—the in-principle feasibility of externalizing tacit knowledge via dialogue—that goes beyond the reader's emphasis on trust and disclosure. The reader's weakest assumption includes 'conversational output can be converted into knowledge that is actually usable by others,' which partially overlaps with my concern. However, I place more weight on the inherent tension between the paper's own definition of tacit knowledge and its proposed language-only capture mechanism, and on the technical misapplication of CoT. These issues do not necessarily invalidate the proposal as a direction, but they strengthen the case that the paper's strongest claim is not merely untested but potentially underspecified. The correct verdict remains UNVERTED—neither accepted as demonstrated nor rejected as impossible—because the paper itself is honest about being early-stage. A concrete empirical test, as described, would be the next step to move from UNVERDICTED toward a substantive evaluation.","tokens_in":7410,"tokens_out":3547,"duration_ms":43300,"concrete_test":"Run a controlled study in a complex, experience-dependent domain (e.g., troubleshooting a manufacturing line). Have N expert operators complete a realistic task while their actions are recorded. Then interview each expert with a SIA using the described LLM+RAG+CoT pipeline to capture their decision process. Independently, have trained novices (or blind raters) attempt to reproduce the expert's decisions from the resulting transcript in a simulated scenario. If the transcript-based performance does not statistically exceed a baseline of existing written documentation (or if raters miss more than a pre-registered threshold of critical decision points), then the conversational externalization fails to capture the tacit component, contradicting the central claim. This directly tests whether the output is usable knowledge, not just fluent prose.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SIAs can preserve and transfer tacit knowledge by engaging employees in empathetic dialogue, using LLMs, RAG, and CoT prompting. The load-bearing assumption is not merely that employees will disclose, but that the very act of conversational externalization can capture what the paper itself defines as 'subjective, experience-based, and difficult to articulate' (Section 2.1). This tension is never resolved: the paper cites Polanyi, whose 'tacit dimension' emphasizes that much knowledge is not fully articulable, yet it proposes a method that is entirely language-based. If tacit knowledge is truly tacit, then a SIA interview can at best elicit explicit-izable fragments (stories, rules, judgments), not the embodied skill or intuitive judgment that constitutes the tacit component. The paper provides no evidence—empirical, analytical, or even a worked example—that such externalization preserves the knowledge's utility for a learner. Section 3's reliance on CoT is also technically suspect: it cites Wei et al. (2022) as if CoT prompting enables the agent to 'ask the right questions in the correct order,' but CoT is a method for eliciting an LLM's own reasoning, not a dialog-management strategy. This misapplication suggests the proposed technical mechanism is not well-grounded. Thus, even granting full trust and disclosure, the core promise—transferring tacit knowledge rather than merely documenting explicit recollections—remains unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes that Socially Interactive Agents (SIAs) can serve as scalable facilitators for preserving and transferring tacit knowledge in organizations. It reviews tacit knowledge management, summarizes SIA research on rapport, trust, and self-disclosure, and outlines a system combining LLMs, RAG, and chain-of-thought prompting to elicit employees' experiential knowledge through empathic dialogue. Application scenarios include onboarding, knowledge retention for retiring experts, and career-biographical development. The paper also discusses ethical and operational risks. It is a vision/position paper: no implementation, empirical data, or formal proof is presented, and the conclusion explicitly acknowledges that the development is at an early stage.","tokens_in":7765,"tokens_out":4565,"duration_ms":51384,"significance":"The paper addresses a practically important problem—the loss of expert knowledge in aging organizations—and proposes a genuinely new role for SIAs as knowledge transfer facilitators. Its strengths are a coherent synthesis of prior SIA findings on trust and disclosure, a clear mapping of organizational pain points, and an explicit treatment of privacy, bias, and transparency concerns. If the proposed mechanism worked, it could complement human facilitators and reduce costs. However, the central value claim is not demonstrated. The step from conversational output to usable tacit knowledge is assumed rather than argued, and the single cited technical mechanism (chain-of-thought prompting) is misapplied. The contribution is therefore a plausible research agenda rather than a validated solution.","major_comments":[{"comment":"The paper defines tacit knowledge as 'difficult to formalize' and 'deeply embedded in the actions, experiences, ideals, values, and feelings of employees,' yet the entire solution is based on natural-language dialogue. No argument is given that conversational elicitation can preserve the utility of the tacit dimension rather than only its explicit-izable fragments. This is load-bearing because the paper's promise is transfer of tacit knowledge, not documentation of recollections. A worked example showing what artifact is produced and how a newcomer would use it to act differently would at least make the claim concrete.","section":"§2.1 and §3"},{"comment":"The statement that 'SIAs could use CoT prompting to ask the right questions in the correct order, just as an experienced interviewer would (Wei et al., 2022)' is a technical misreading. Chain-of-thought prompting elicits intermediate reasoning from a large language model for a given prompt; it is not a dialog-management strategy for selecting questions in a conversation. Since this is the paper's only concrete mechanism for steering knowledge elicitation, the technical foundation is not sound. The authors should either correct this claim and cite appropriate dialogue-policy or active-learning methods, or specify an architecture that makes question selection explicit.","section":"§3"},{"comment":"The trust-and-disclosure evidence comes from health screening and job-interview training, where the incentives, stakes, and power relations differ from an employee interacting with an organization-provided AI. The paper asserts that employees are reluctant to share tacit knowledge and that trust-building will overcome this, but it offers no study, pilot, or even a scenario transcript to support transferability to this new context. Because the system's entire value depends on employees disclosing substantive experiential knowledge, this is a load-bearing assumption that needs at least a formative user study or a clearly testable design hypothesis.","section":"§3–4"},{"comment":"The application scenarios claim that the approach 'could significantly shorten the period of onboarding' and provide a 'digital legacy' for retiring experts, but no comparison with human-facilitated knowledge transfer is provided. The scalability argument assumes that SIA-mediated elicitation is at least comparable in effectiveness to human facilitation; without a baseline, the central value proposition is unsubstantiated. A task analysis, cost model, or small-scale evaluation would make the proposal more than an analogy.","section":"§4"}],"minor_comments":[{"comment":"Figure 1 and Figure 2 captions appear, but the figures themselves are not present in the manuscript; the Figure 1 caption also appears truncated ('Vision of an AI preservation'). Please provide the figures and complete captions.","section":"Figures"},{"comment":"The reference list is not consistently formatted: for example, the Beyrodt et al. entry is not a standard citation, some entries lack page ranges, and one URL uses an unsecured 'http'. Please normalize to the journal's reference style.","section":"References"},{"comment":"The claim that 'there are no known systems' integrating SIA-based and immersive technologies into knowledge transfer is overly strong without a systematic literature search. A softer wording such as 'to our knowledge, no published systems...' would be more appropriate, together with a brief search scope.","section":"§6"},{"comment":"The example of tacit knowledge in 'manual labor' is unnecessarily narrow; the subsequent examples of decision-making and creativity already make the point. Replacing 'manual labor' with 'sensorimotor skills' would better align with the knowledge-society framing.","section":"§2.1"}],"recommendation":"major_revision","confidential_remarks":"This is a position/vision paper rather than an empirical study. If the journal publishes such contributions, the revision path is feasible: the tacit-knowledge externalization gap, the CoT misreading, and the missing evidence for disclosure transferability must be addressed. If the journal's scope requires empirical validation, then the manuscript is currently out of scope. The CoT error in particular should be corrected regardless of the venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a conceptual proposal, not a study, and it reads like one. What is actually new is the application of socially interactive agents (not just chatbots) to organizational tacit knowledge transfer, framed around the facilitator role and grounded in the SIA literature on rapport and reduced disclosure barriers. The paper does that well: it spells out why traditional methods are resource-heavy, identifies trust as the decisive variable, and honestly says no system exists yet. The survey of relevant prior SIA work is legitimate, including the self-citations, which are all on-topic. Credit where due: this is a usable roadmap for a plausible research direction.\n\nThe soft spots are real, and one is load-bearing. Section 2.1 defines tacit knowledge as 'subjective, experience-based, and difficult to articulate' and cites Polanyi, whose whole point is that much knowledge resists articulation. The solution approach, however, is entirely language-based: empathic dialogue, LLM generation, RAG, CoT. The paper never squares that circle. At best, dialogue can elicit explicit-izable fragments—stories, rules, judgments—but the embodied, intuitive component may not survive the transfer. That does not kill the idea, but it needs to be acknowledged and argued, not skipped. The paper currently treats externalization through conversation as unproblematic.\n\nSecond, the CoT citation is wrong. The authors cite Wei et al. (2022) to support the claim that CoT lets the agent 'ask the right questions in the correct order,' but chain-of-thought prompting is a method for eliciting an LLM's own reasoning over a prompt—not a dialogue-management strategy. That is a technical misreading, and it matters because it is the sole justification for how the agent steers the conversation. Easy to fix, but it should be fixed.\n\nThere is no empirical data, no implementation, no worked example. For a position paper that is acceptable, but the authors occasionally drift into 'the potential is enormous' territory, and the application scenarios are speculative. The paper also concedes the ethical and privacy risks, which is good, but these are listed rather than analyzed.\n\nThe central argument does not collapse; it just is not established. The paper is a proposal with a promising premise and correctable flaws. Who gets value: researchers working on SIAs, knowledge management, or AI-mediated organizational learning—especially anyone designing studies around knowledge preservation. I would not cite it as evidence, but I might cite it as a formulation of a research gap.\n\nEngage with it. Send it to peer review; the concept is relevant and the issues are discussable rather than fatal. The authors should be pushed on the tacit-externalization tension and the CoT citation before publication.","headline":"A clearly written, honest position paper proposing SIAs as tacit-knowledge transfer facilitators—but the core mechanism is in tension with its own definition of tacit knowledge, and the CoT citation is conceptually off.","tokens_in":8205,"tokens_out":1652,"would_cite":false,"duration_ms":21330,"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":"A position paper argues that Socially Interactive Agents—AI avatars that converse through natural language and social cues—can act as scalable knowledge-transfer facilitators, using LLM dialogue, retrieval-augmented generation, and chain-of","keywords":["tacit knowledge","knowledge transfer","socially interactive agents","LLM dialogue","retrieval-augmented generation","chain-of-thought prompting","onboarding","knowledge retention"],"falsifier":"A controlled comparison in which senior employees approaching retirement are interviewed by the proposed SIA and by a human facilitator, with recorded answers rated by successors for undocumented, actionable heuristics. If the SIA elicits no more or worse usable know-how than a simple written questionnaire, or if employees refuse to discuss the reasoning behind real past decisions, the central claim is contradicted.","tokens_in":7364,"feed_emoji":"🧠","tokens_out":6789,"duration_ms":74230,"temperature":0.7,"pith_summary":"The paper argues that the hardest part of organizational knowledge retention—capturing the unwritten, experience-based know-how of employees—can be tackled by Socially Interactive Agents acting as AI-driven knowledge-transfer facilitators. It claims these agents can replace or scale the human facilitator role by combining a socially embodied presence with large language model dialogue, retrieval-augmented generation, and chain-of-thought prompting, so that employees externalize insights they would not normally put into words. If this works, organizations could systematically interview departing experts, preserve the heuristics behind their decisions, and shorten onboarding for new hires. The paper frames this as a new direction because existing methods such as documentation, mentoring, and human-facilitated interviews are effective but do not scale, while current AI systems mostly capture explicit knowledge rather than deep experiential knowledge. It does not report an empirical test; it lays out the requirement that employees trust the agent enough to disclose, and sketches the technical architecture and application scenarios that would realize the idea.","feed_headline":"Social AI agents offer a way to preserve retiring experts' know-how","feed_subtitle":"A proposal pairs conversational agents with LLMs, RAG, and chain-of-thought prompts to turn expert heuristics into company knowledge.","key_machinery":"The central object is the Socially Interactive Agent as a knowledge-transfer facilitator: a virtually or physically embodied agent that communicates through natural multimodal behavior—verbal, paraverbal, and nonverbal—just as a human conversation partner would. The mechanism that carries the argument is the combination of three technologies inside that social frame: large language models generate contextually relevant dialogue; retrieval-augmented generation connects the conversation in real time to organizational databases, guidelines, and documented experiences; and chain-of-thought prompting gives the agent an interview strategy, letting it ask probing questions in a structured order tha","core_discovery":"The paper's central claim is that established findings about Socially Interactive Agents—people treat them as social partners, build rapport with them, and sometimes disclose more sensitive information to them than to humans—can be redirected to knowledge management. It proposes a system in which an SIA conducts long-term, empathic dialogues with employees, using large language models to generate natural conversation, retrieval-augmented generation to pull in relevant company documents and prior recorded experiences, and chain-of-thought prompting to steer the conversation toward the reflective 'how' of expert decision-making. The intended result is a recorded, context-linked account of heur","pith_inferences":["A natural test of the proposal is to compare the quantity and quality of heuristics elicited by a social agent versus a human facilitator under blind review; if disclosure is comparable, the scalability argument holds, and if it is not, the social presence claim is weakened.","The approach could be extended beyond retrospective exit interviews to in-situ capture: if the agent is present during daily work, it might record tacit knowledge as it is exercised rather than relying on later recall, reducing the distortion that memory introduces.","A testable design question the paper leaves open is how much embodiment is needed: measuring willingness to disclose with a voice-only interface versus an embodied agent would reveal whether the social machine or the conversational strategy is the true driver of elicitation.","Because the paper treats trust as the load-bearing variable, an honorable falsification path is to measure whether employees with privacy concerns withhold the reasoning behind actual past decisions even when they state a preference for the agent over a human interviewer."],"forward_implications":["Knowledge retention could scale beyond one-on-one human-facilitated interviews: every retiring expert, not just a select few, could receive a structured exit interview that captures decision heuristics.","Captured knowledge would be context-linked rather than stored as isolated text: RAG could tie an employee's statements to the actual process documents and prior examples, making the knowledge usable by successors.","Onboarding could become personalized and always available, with an agent that guides new hires through processes, personnel, and training, potentially reducing time to productivity.","Chain-of-thought-guided interviewing could shift knowledge capture from facts and figures to the 'how'—judgments, assumptions, and unwritten rules—addressing a known gap in documentation and training.","Trust, transparency, and personalization would become acceptance criteria for any real deployment, since the paper's own analysis concludes that without them the agent will not be used."],"supporting_citations":[{"why":"Defines tacit knowledge as deeply personal and hard to articulate, establishing the target phenomenon the approach is built to externalize.","marker":"Polanyi, 1966"},{"why":"Provides the knowledge-creation and transfer foundation that frames tacit knowledge as a core organizational resource.","marker":"Nonaka & Takeuchi, 1995"},{"why":"Documents the 'leaving expert' problem and the human knowledge-transfer facilitator role that the proposed SIAs are meant to scale.","marker":"Hofer-Alfeis, 2008"},{"why":"Supplies the definition and research basis for Socially Interactive Agents as multimodal, socially intelligent interactants.","marker":"Lugrin, 2021"},{"why":"Shows that virtual humans increase willingness to disclose sensitive information, the key empirical premise for expecting employees to open up to an SIA.","marker":"Lucas et al., 2014"},{"why":"Establishes warmth and competence as the dimensions that build trust, which the agent design uses to overcome reluctance to share.","marker":"Fiske, Cuddy, & Glick, 2007"},{"why":"Introduces chain-of-thought prompting, the mechanism proposed to guide structured, reflective questioning in dialogue.","marker":"Wei et al., 2022"},{"why":"Surveys retrieval-augmented generation, the mechanism proposed to connect employee statements to organizational knowledge in real time.","marker":"Gao et al., 2023"}],"fun_headline_variants":["AI agents coax tacit know-how from retiring experts","Conversational AI captures experts' unspoken wisdom","Social AI preserves hard-to-articulate expertise","LLM-driven agents turn expert heuristics into company knowledge","Talking AI agents extract tacit knowledge before it retires"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The approach works only if employees actually tell the agent their deep, experience-based knowledge—and if what they say can be turned into guidance a successor can use; the paper asserts that trust-building design will achieve this, but does not test it.","fun_headline_variants_meta":{"raw":{"variants":["AI agents coax tacit know-how from retiring experts","Conversational AI captures experts' unspoken wisdom","Social AI preserves hard-to-articulate expertise","LLM-driven agents turn expert heuristics into company knowledge","Talking AI agents extract tacit knowledge before it retires"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000465,"raw_usage":{"total_tokens":2178,"prompt_tokens":786,"completion_tokens":1392,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":1317}},"tokens_in":530,"tokens_out":1392,"duration_ms":11487,"temperature":1.0,"reasoning_tokens":1317,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:19:47.409925+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled comparison in which senior employees approaching retirement are interviewed by the proposed SIA and by a human facilitator, with recorded answers rated by successors for undocumented, actionable heuristics. If the SIA elicits no more or worse usable know-how than a simple written questionnaire, or if employees refuse to discuss the reasoning behind real past decisions, the central claim is contradicted.","supporting_citations":[],"review_version":1}