{"id":"9073e953-2241-40f4-85cb-7cf0e2c0a169","arxiv_id":"2501.00383","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"Inner Thoughts, a framework that gives conversational AI a continuous stream of covert thoughts with self-evaluated motivation, was rated more natural and coherent than a next-speaker-prediction baseline in simulated group chats.","lead":"This paper proposes a framework that lets conversational AI generate a private train of 'inner thoughts' during group chats and use them to decide when to speak, instead of just predicting who talks next. It reports that human raters preferred conversations produced by this approach over a next-speaker-prediction baseline 82% of the time in simulated group chats.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 82% preference and across-the-board significance likely confound the Inner Thoughts mechanism with LLM capability: the baseline is fixed to fine-tuned GPT-3.5 while condition 2's backend is unspecified and probably stronger; no matched-model ablation exists.","rationale":"Good-faith reading: The paper contributes a well-motivated framework, a think-aloud study with 24 participants, two open-sourced systems, and a plausible design space. The central claim, however, is comparative: 'significantly surpasses existing baselines.' For that claim to hold, the comparison must isolate the Inner Thoughts mechanism. The weakest point in the argument is the asymmetry between the two arms: the baseline is pinned to fine-tuned GPT-3.5, while the Inner Thoughts condition is not tied to any stated backend in Section 6.2. Because the framework makes several LLM calls per trigger, it is likely running a much stronger model, and the reported effect sizes (e.g., turn appropriateness U=577, p=2.4e-6; 82% preference) are exactly what one would expect from a generation-quality gap. The reader's assessment identified baseline fairness and missing ablations; my check sharpens this to a specific model-capability confound and to the absence of any within-framework ablation separating the thought reservoir and motivation scoring from the raw LLM. The Section 7 user study does not include the baseline, so it cannot rescue the comparative claim. This is fixable: a matched-backend replication and an ablation study would determine whether the inner-thought mechanism itself is responsible. Therefore the CONDITIONAL verdict should remain.","tokens_in":25172,"tokens_out":4602,"duration_ms":42989,"concrete_test":"Re-run the Section 6 simulation with both conditions using the same LLM backend (e.g., GPT-4o for both) and equalize the number of LLM calls per trigger, so the baseline also uses the same response-generation model as Inner Thoughts; then compare the seven metrics. If the significant differences shrink or vanish, the central claim is unsupported and a new controlled baseline is needed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 6.2 specifies the baseline exactly: 'We used the fine-tuned GPT-3.5 model we evaluated in section 3 in condition 1 to predict the next speaker, and prompt the model to generate responses based on its persona.' Condition 2 is described only as 'the framework described in section 5,' with no LLM backend stated. The Inner Thoughts pipeline makes multiple LLM calls per trigger (utterance interpretation, system-1/system-2 thought formation, G-Eval-style scoring, tone restyling), and the playground/app elsewhere uses contemporary models; Table 1 tests GPT-4-turbo/GPT-4o. If condition 2 runs on GPT-4-class models, the measured superiority on anthropomorphism, coherence, intelligence, and turn-taking—and the 82% preference—may reflect generation quality rather than the inner-thought mechanism. This is not merely a missing robustness check: the paper's own Section 5 states 'We recognize conducting formal ablation studies an important direction for future work,' and the Section 7 user study compares three Inner Thoughts proactivity settings, not Inner Thoughts vs. a baseline. Thus the headline comparative claim rests on a single simulated comparison in which the two arms are not matched for model capability, compute, or number of LLM calls. The Appendix also lists 'Overt proactivity = 3.95, Covert proactivity = 0.1' (Section 6.3) where the labels appear swapped relative to the definitions in Section 5.5, a small but telling sign that the simulation parameters need auditing. The concern is not that the framework is valueless; it is that this evaluation cannot yet separate the mechanism from the engine.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Inner Thoughts framework, in which a conversational AI continuously generates a covert stream of thoughts in parallel to an ongoing multi-party conversation, retrieves relevant long-term memories, evaluates each thought's 'intrinsic motivation' to be expressed using LLM-based scoring over heuristics derived from a 24-participant think-aloud study, and then decides when and how to participate on the basis of thresholds and turn-taking type. The authors instantiate this framework in a playground web app and a Slackbot, report a technical evaluation of simulated conversations against a next-speaker-prediction-plus-persona baseline, and report a user study comparing three proactivity settings of the framework.","tokens_in":25585,"tokens_out":2693,"duration_ms":28840,"significance":"The paper addresses an important and underexplored problem: how a conversational agent should decide to self-select in multi-party social conversations, where next-speaker prediction is known to be difficult. The think-aloud study and the resulting ten heuristics are a useful qualitative contribution, and the two open-sourced, deployed implementations make the framework concrete and reproducible. The human raters' preference for the Inner Thoughts conversations over the baseline in the simulated evaluation is encouraging evidence for the approach. However, the central comparative claim that the framework 'significantly surpasses existing baselines' currently rests on a single simulated comparison in which the two arms are not matched for model capability, number of LLM calls, or parameter tuning. The paper is therefore best read as presenting a promising framework with an initial proof-of-concept evaluation rather than a fully established superiority claim.","major_comments":[{"comment":"The baseline condition is specified as a fine-tuned GPT-3.5 model for next-speaker prediction with persona-based response generation, while the Inner Thoughts condition is described only as 'the framework described in section 5' with no LLM backend stated. If condition 2 runs on a stronger model (e.g., GPT-4-class) or makes many more LLM calls per turn, then the observed advantages in anthropomorphism, coherence, intelligence, and turn-taking appropriateness, as well as the 82% preference, could be driven by generation quality or prompt engineering rather than by the inner-thought mechanism. This is a load-bearing point for the paper's headline claim; the authors should either report the exact backend used for condition 2 and run a matched-model ablation, or temper the superiority claim accordingly.","section":"§6.2, Table 2, Figure 7"},{"comment":"The simulation parameters listed in §6.3 ('Overt proactivity = 3.95, Covert proactivity = 0.1') are inconsistent with the definitions in §5.5, where overt proactivity is the system1Prob parameter (0-1) and covert proactivity is the imThreshold parameter (1-5). The values appear to be swapped, and 3.95 cannot be a valid system1Prob. This suggests that the simulation configuration needs an audit. Since participation behavior depends directly on these thresholds and the authors state in §5 that the hyperparameters are chosen empirically, the paper should report the exact parameter settings used in the simulations and ideally include a sensitivity analysis.","section":"§6.3 and §5.5"},{"comment":"The thought-evaluation stage uses an LLM to score each thought's intrinsic motivation with criteria and five-level definitions that were themselves derived from a think-aloud study, and the same class of LLM also generates the thoughts. This creates a partial self-consistency loop: the model's judgments about its own thoughts may be coherent without corresponding to human judgments of appropriate participation. Although the human preference data in §6.6 provide independent evidence, the paper does not report whether human raters agree with the LLM's intrinsic-motivation scores. A small validation study comparing LLM motivation scores with human ratings of the same thoughts would substantially strengthen this component.","section":"§5.4 and §6.6"},{"comment":"The user study in §7 compares three proactivity settings of the Inner Thoughts framework (Non-stop Chatter, Active Contributor, Selective Participant) and does not include the next-speaker-prediction baseline. Consequently, the abstract's statement that the framework 'significantly surpasses existing baselines' is supported only by the simulated evaluation in §6. The authors should either add a baseline condition to the user study or revise the abstract and introduction to specify that the superiority claim is based on the technical evaluation of simulated conversations, while the user study demonstrates that different proactivity settings are perceptible and that the Active Contributor setting is preferred over the other two.","section":"§7 and Abstract"}],"minor_comments":[{"comment":"The word 'calculaed' appears in the sentence describing the score formula; it should be 'calculated'.","section":"§5.4"},{"comment":"The stacked bar plot in Figure 7 would benefit from error bars or a point plot showing the distribution of ratings, because the Mann-Whitney U tests are reported on the underlying Likert responses but the figure does not convey variance.","section":"§6.6 and Figure 7"},{"comment":"The identification accuracy rates in §7.3.2 are reported against a 33.3% chance baseline, but with 12 participants the confidence intervals are wide; the authors should be careful in describing these as evidence that users can reliably distinguish the three styles.","section":"§7.2"},{"comment":"The notation 'd_p = lambda(t - tau_p)' is introduced twice in the same section; once for memory decay and once for the silence-based motivation increase, with different decay-rate values. Using distinct symbols or names would avoid confusion.","section":"§5.4"}],"recommendation":"major_revision","confidential_remarks":"The central idea is interesting and likely of interest to the CHI audience, but the main comparative claim needs stronger evidence than the current single simulation with an unmatched baseline. I do not see evidence of intentional misreporting; the inconsistencies in the proactivity parameter values and the omission of the condition-2 LLM backend are most plausibly presentation oversights, but they are consequential because they directly affect the evaluation's validity. A matched-model ablation or a clear statement of the backend with a same-LLM control would be the minimum needed to support the abstract's wording."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper. The Inner Thoughts framework is a real idea: running a covert thought stream in parallel to the dialogue, scoring each thought on intrinsic motivation, and letting that score drive whether and when to speak is genuinely different from next-speaker prediction. The formative think-aloud study with 24 participants is solid work and the derived heuristics are useful. The five-stage pipeline is clearly described, the playground and Slackbot are open-sourced, and the Section 7 user study, while baseline-free, does show that different proactivity settings are perceptible and that a moderate setting is preferred.\n\nThe problem is the headline comparative claim. The stress-test note is correct: condition 1 is explicitly a fine-tuned GPT-3.5 baseline, condition 2's LLM backend is never stated, and the Inner Thoughts pipeline makes substantially more LLM calls per trigger. With every metric significantly higher and an 82% preference in that single simulated comparison, you cannot separate the mechanism from the engine. The paper itself admits in Section 5 that formal ablation studies are left to future work, so the abstract's 'significantly surpasses existing baselines' is not supported by the evidence as it stands. The Section 7 user study compares three proactivity settings, not Inner Thoughts against next-speaker prediction, so it doesn't rescue the claim. Also, the parameter list in Section 6.3 appears to have the overt/covert labels swapped relative to the definitions in Section 5.5--another sign that the simulation parameters need auditing.\n\nNone of this kills the framework. The qualitative findings, the heuristics, and the open-source systems are worth having, and the authors are candid about hand-tuned thresholds in Section 8.4. But the central evaluation needs to be redone with matched model backends, controlled compute, and ablations that isolate the thought reservoir and motivation scoring. As it stands, I'd treat the comparative claim as unproven rather than false.\n\nThis paper is for HCI and conversational AI researchers working on proactive agents. It deserves a serious referee; I'd send it to peer review with a clear request for major revision. I would engage with the framework and cite it for the system and qualitative results, but not for the comparative result.","headline":"A solid framework with an honest but unproven comparative claim: the evaluation doesn't yet separate the inner-thought mechanism from the underlying LLM capability.","tokens_in":26058,"tokens_out":2469,"would_cite":true,"duration_ms":25542,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An AI that continuously forms and scores its own inner thoughts can join group conversations proactively, outperforming next-speaker prediction on anthropomorphism, coherence, intelligence, and turn-taking appropriateness.","keywords":["conversational agent","multi-party conversation","proactive AI","inner thoughts","intrinsic motivation","turn-taking","large language models","mixed-initiative interaction"],"falsifier":"A controlled comparison that keeps the response-generation step identical and varies only the participation decision—for example, replacing the intrinsic-motivation gate with random thought selection at the same speaking rate—would settle whether the thought-evaluation mechanism carries the effect; the paper reports no such ablation.","tokens_in":24989,"feed_emoji":"🧠","tokens_out":4613,"duration_ms":41835,"temperature":0.7,"pith_summary":"The paper argues that the standard way of making AI proactive in multi-party chats—predicting who will speak next and then generating a persona-based response—cannot produce natural participation, because in self-selection moments no observable cue determines the next speaker. Its alternative is to give the AI a continuous, covert stream of inner thoughts that runs alongside the conversation, each thought scored by how strongly the AI is intrinsically motivated to say it. The authors derive eight motivation heuristics from a think-aloud study with 24 participants, encode them as LLM evaluation criteria, and build the five-stage Inner Thoughts framework around them. In simulated conversations judged by humans, the framework outperforms the next-speaker-prediction baseline on all seven rated dimensions, and participants preferred it 82% of the time. The claim is that modelling the inner impulse to speak, rather than the outer turn-taking signal, is what makes conversational AI feel proactive.","feed_headline":"Inner thoughts let AI join group chats naturally","feed_subtitle":"Modeling an AI's motivation to speak beats next-speaker prediction on all seven quality metrics.","key_machinery":"The central object is the thought reservoir: a continuously updated set of covert candidate contributions that the AI forms in parallel with the overt conversation. Each cycle runs five stages—trigger, retrieval, thought formation, evaluation, and participation—where an incoming message or a ten-second pause triggers retrieval of relevant long-term-memory items by embedding saliency, formation of one fast 'System 1' and two deliberate 'System 2' thoughts per batch, and an LLM evaluation that assigns each thought a 1–5 intrinsic-motivation score, adjusted upward the longer the AI has been silent. Participation is then gated by interpretable thresholds: an imThreshold for open turns, system1Prob for low-motivation fill-in responses, and an interruptThreshold that lets the AI override turn allocation when a thought is urgent. This machinery is what carries the argument, because it replaces 'who is externally most likely to speak next' with 'which internal thought is most worth saying now'.","core_discovery":"On the paper's own terms, the discovery is that AI proactivity in multi-party conversation is better modelled as an internal, motivation-driven process than as a prediction problem. The paper shows that even strong LLMs predict the next speaker at near-chance levels in self-selection cases, and that relabelling those cases as 'anyone' improves the task but sidesteps the real question of who should actually contribute. The Inner Thoughts framework answers that question by generating thoughts in parallel, scoring each thought on relevance, information gap, expected impact, urgency, coherence, originality, balance, and dynamics, and letting a threshold decide whether to speak, remain silent, or even interrupt an allocated turn. In the technical evaluation, all seven metrics—anthropomorphism, coherence, engagement, intelligence, turn appropriateness, initiative, and adaptability—were rated significantly higher for Inner Thoughts conversations than for the baseline, with a stated 82% preference for the Inner Thoughts condition.","pith_inferences":["The 82% preference should be read with caution: the baseline's response generation was tied to fixed personas, so part of the gap may come from weaker content rather than from the thought-gating mechanism itself.","A testable extension would hold the response-generation step fixed and vary only the participation decision, isolating the contribution of intrinsic-motivation scoring.","If the thought reservoir were surfaced to users, it could act as a transparency interface, letting people see and veto the AI's unspoken reasoning—a direction the paper raises but does not test.","The near-chance self-selection results suggest a domain ceiling for history-based turn-taking models; any gains from richer multimodal cues may still need an internal motivation model to convert attention into appropriate speech."],"forward_implications":["Group-chat assistants could engage without being named, contributing when a thought clears the motivation bar.","Next-speaker prediction may be better reframed as a two-level task: detect whether a turn is open or allocated, then let content motivation decide participation.","Adjustable proactivity parameters give users direct control over how talkative, assertive, or reserved an AI is.","Because the evaluation criteria are explicit and score-based, the same machinery can double as an automatic, explainable benchmark for proactive dialogue quality.","The framework's scope extends to task-oriented settings such as brainstorming and negotiation by swapping in goal-aligned evaluation criteria."],"supporting_citations":[{"why":"Supplies the next-speaker prediction task and prediction window that the paper evaluates and criticises.","marker":"[15]"},{"why":"Another transformer-based turn-taking model that represents the prediction approach the paper argues is insufficient.","marker":"[20]"},{"why":"Multi-party chat agents baseline using next-speaker prediction and persona responses; the paper's direct comparison point.","marker":"[63]"},{"why":"MPC corpus provides the labelled multi-party chat data used to test next-speaker prediction in self-selection versus turn-allocation cases.","marker":"[53]"},{"why":"PersonaChat seeds the agent personas and defines the persona-based response generation used in the baseline condition.","marker":"[71]"},{"why":"Sacks et al.'s Simplest Systematics supplies the turn-allocation versus self-selection distinction that motivates the whole approach.","marker":"[52]"},{"why":"Dual-processing theory grounds the System 1 and System 2 thought formation split in the framework.","marker":"[21]"},{"why":"G-Eval inspires the weighted chain-of-thought evaluation pipeline used for intrinsic-motivation scoring.","marker":"[38]"},{"why":"Retrospective think-aloud is the method used in the formative study that produced the eight motivation heuristics.","marker":"[26]"}],"fun_headline_variants":["AI with inner thoughts speaks up at the right moments","Proactive AI models its motivation to talk, not just cues","Inner Thoughts lets AI choose when to join group chats","Motivation-driven AI beats next-speaker prediction in chats","AI that thinks before speaking joins conversations naturally"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the next-speaker-prediction-plus-persona system is a fair and representative baseline; if it is weaker than typical proactive systems, the reported advantage may come from better response generation or extra LLM calls rather than from the inner-thought mechanism.","fun_headline_variants_meta":{"raw":{"variants":["AI with inner thoughts speaks up at the right moments","Proactive AI models its motivation to talk, not just cues","Inner Thoughts lets AI choose when to join group chats","Motivation-driven AI beats next-speaker prediction in chats","AI that thinks before speaking joins conversations naturally"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000186,"raw_usage":{"total_tokens":1316,"prompt_tokens":924,"completion_tokens":392,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":315}},"tokens_in":540,"tokens_out":392,"duration_ms":4546,"temperature":1.0,"reasoning_tokens":315,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:52:06.691812+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled comparison that keeps the response-generation step identical and varies only the participation decision—for example, replacing the intrinsic-motivation gate with random thought selection at the same speaking rate—would settle whether the thought-evaluation mechanism carries the effect; the paper reports no such ablation.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MPC corpus provides the labelled multi-party chat data used to test next-speaker prediction in self-selection versus turn-allocation cases."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Dual-processing theory grounds the System 1 and System 2 thought formation split in the framework."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Retrospective think-aloud is the method used in the formative study that produced the eight motivation heuristics."}],"review_version":1}