{"id":"3bbac483-c724-4ba1-8ceb-9d9448b677ec","arxiv_id":"2501.00953","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A literature review that identifies a lack of incremental dialogue management research and proposes desiderata for robots that converse with people word by word.","lead":"This survey reviews twenty-five years of research on incremental spoken dialogue systems and finds that almost no work has been done on the dialogue manager, the component that decides when and what a robot should do. It then offers requirements for building incremental dialogue managers for robots, which matters as language models are increasingly integrated into embodied systems.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'very little research on incremental DM' claim may be an artifact of classifying turn-taking and timing work outside DM, even though Section 3.1.3 assigns timing to DM.","rationale":"The reader's weakest_assumption is exactly the definitional division of labor: if timing and revision are handled outside DM, the scarcity claim is an artifact. I agree with this, and I find it to be the most load-bearing concern because the paper itself creates the tension. Section 3.1.3 says timing is part of DM, yet the incremental-DM survey in Section 3.2 does not treat the turn-taking and timing literature as DM work. This is not a dispute about external consensus; it is an internal consistency issue that determines whether the central contribution is a genuine gap or a re-labeling of existing work. The paper is otherwise a thoughtful review with useful desiderata, and the claimed gap may well be real if the strict definition of DM excludes timing-only systems. But the current presentation makes the finding sensitive to classification choices. Since the reader's verdict is already CONDITIONAL and this concern supports that conditional status rather than overturning it, no adjustment is needed. The concrete test I propose would settle the matter by counting how many works satisfy the paper's own functional definition, making the classification explicit and reproducible.","tokens_in":26191,"tokens_out":3510,"duration_ms":34523,"concrete_test":"Build a rubric from the paper's own definitions in Sections 3.1 and 4.1: a system counts as incremental DM if it (a) updates a state or decision from partial input, (b) selects an action or a timing decision, and (c) supports revision or re-decision after new input. Apply this rubric to all works cited in Section 3.1.3 and the replanning references in Section 3.1.2, plus an ACL Anthology search for 'incremental' combined with 'dialogue policy', 'turn-taking', 'response timing', and 'dialogue management', excluding the authors' own papers. If at least ten independent works satisfy the rubric, the headline claim should be revised from 'very little research on incremental DM' to 'few systems explicitly label themselves incremental DMs'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that there is very little research on incremental dialogue management, in contrast to incremental ASR, NLU, and NLG. Section 3.1 defines DM as state tracking plus action/response selection, and Section 3.1.3 explicitly says DM is responsible not only for what to do but also when to do it, locating timing inside DM. Under that functional definition, several works the paper itself reviews in Section 3.1.3 appear to be incremental decision-making systems: Raux and Eskenazi (2009) decide whether to grab, release, wait for, or keep the floor; end-of-turn prediction models decide when a response can begin; Roddy and Harte (2020) generate response timings; and voice activity projection methods decide when the system should start speaking. If these count as incremental DM, the scarcity finding weakens considerably. The paper also mentions robot replanning (Cashmore et al., 2019; Garrett et al., 2020; Zhou et al., 2023) as related but does not survey that literature as incremental action management, even though replanning is arguably incremental decision-making for embodied dialogue agents. The load-bearing assumption is therefore not about what the literature contains, but about how it is classified. The paper's own definitional commitments make this classification vulnerable: either timing is part of DM and Section 3.1.3 contains many incremental DM contributions, or timing is not part of DM and the desideratum in Section 4.1 that 'Incremental DM is responsible for timing' is outside the surveyed definition. Both options put pressure on the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper surveys incremental spoken dialogue systems with a focus on dialogue management (DM) and argues that, unlike ASR, NLU, and NLG, incremental DM has received very little research attention. The authors motivate incrementality by the needs of human-robot interaction, explain frameworks such as Incremental Units and the restart vs. update distinction, review incremental modules and full systems, and propose desiderata for incremental DM on robotic platforms, including responsibility for timing, acting on incomplete information, and making fast, small decisions concurrently.","tokens_in":26444,"tokens_out":3991,"duration_ms":38574,"significance":"If the central scarcity claim is robust, the paper identifies a genuinely under-explored area with practical importance: incremental decision-making for dialogue and robot action management in the age of LLMs. The paper is useful as a synthesis: it collects a broad set of references, clearly explains the IU framework and the restart/update distinction, connects incremental processing to non-monotonicity and replanning, and offers concrete desiderata that could guide future work. The authors are appropriately careful in places, phrasing the finding as 'very little research' rather than 'none'. However, the main contribution is the gap claim itself, and that claim currently rests on an implicit and internally inconsistent boundary between DM and adjacent modules, so the paper's significance depends on the strength of that boundary.","major_comments":[{"comment":"The paper's scope for DM is internally inconsistent. Section 3.1.3 explicitly states that DM is about not only what to do but also when to do it, and Section 4.1 lists 'Incremental DM is responsible for timing' as a desideratum. Under this definition, several systems reviewed in Section 3.1.3—Raux and Eskenazi (2009) deciding whether to grab, release, wait for, or keep the floor; end-of-turn prediction models; Roddy and Harte (2020) generating response timings; and voice activity projection methods—are incremental decision-making contributions. Yet Section 3.2 excludes these works from its count of incremental DM research and concludes that very little such research exists. The scarcity claim is therefore an artifact of the classification unless the paper either includes timing-focused work as incremental DM or explicitly justifies a narrower definition that excludes timing.","section":"§3.1.3, §3.2, §4.1"},{"comment":"The central claim is a quantitative statement about the literature, but the review does not report a systematic selection method, search protocol, inclusion criteria, or corpus size. The abstract states 'We find that there is very little research on incremental dialogue management,' which reads as an empirical finding, yet no procedure is described for identifying the set of works surveyed. The claim is not independently verifiable and could reflect the authors' selection rather than the actual state of the field. The authors should either soften the claim to the set of works they examined or add an explicit methodology and inclusion criteria.","section":"§1, §3.2"},{"comment":"The title and motivation emphasize robot action management, but the review does not survey the robotics replanning literature that it identifies as relevant. Section 2.2.1 and Section 3.1.2 mention replanning (Cashmore et al., 2019; Garrett et al., 2020; Zhou et al., 2023) as an analogue to revising DM decisions, and the conclusion argues that incremental decision-making is needed for robots, yet no incremental action management or replanning work is actually reviewed. Without such coverage, the paper cannot support its implicit claim that incremental decision-making in embodied settings is under-researched; the gap may be a result of excluding this literature rather than a property of the field.","section":"§2.2.1, §3.1.2, §5"}],"minor_comments":[{"comment":"There is a typo in 'people antrhopomorphize robots'—should be 'anthropomorphize'.","section":"§1"},{"comment":"The phrase 'as the first word is uttered is uttered' contains a duplicated 'is uttered' and should be corrected.","section":"§2.1"},{"comment":"The sentence 'Ghigi et al. (2014b) showed that that incremental dialogue strategy' has a doubled 'that'.","section":"§2.3.4"},{"comment":"The text 'V oice Activity Detection' has unusual spacing in 'Voice Activity Detection'; this appears to be a formatting artifact and should be fixed.","section":"§4.1"},{"comment":"The figure caption includes a raw URL instead of a formal citation to the Localized Narrative dataset paper; a proper reference would be more appropriate.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a genuinely useful synthesis, and the desiderata are valuable. The main risk is that the central 'very little research' claim is not yet robust because the scope of 'incremental DM' is drawn inconsistently and the selection method is absent. I would not reject it, as the issues are addressable: the authors can clarify their definition, report how the literature was chosen, and either incorporate timing-focused work or explicitly exclude it. I would also gently note that the authors cite their own previous work very frequently; this is normal in a niche field, but a more systematic selection would help balance the presentation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful review, and the gap it names is real, but the way it states the gap is vulnerable to its own definition of dialogue management. The restart-versus-update framing is the most useful thing in it, and the desiderata are practical. But the central finding—'very little research on incremental dialogue management'—depends on a boundary that the paper itself blurs.\n\nWhat is new: it is not new to say ASR, NLU, and NLG have been incrementalized, but the paper does a service by putting DM at the center and asking what incrementality means for the decision-making component. The restart/update distinction is a clean lens, and applying it to DM makes the open problem concrete. The discussion of LLM monotonicity is accurate and well-phrased.\n\nThe soft spot is classification. Section 3.1.3 explicitly says DM is responsible for when to act, then surveys turn-taking and end-of-turn prediction work under that heading. But Section 3.2, which reviews 'incrementalizing DM', does not count that same work as incremental DM. Raux and Eskenazi's finite-state turn-taking model, the end-of-turn predictors, Roddy and Harte's response timing model, and the VAP work are all making timing decisions under the paper's own definition. If those count, the scarcity claim is much weaker. If they do not count, the paper needs to say why, and then the desideratum that 'incremental DM is responsible for timing' is sitting outside the surveyed definition. Either way, the central claim needs qualification. The robot replanning literature is another missing piece; it is cited but not surveyed, and replanning is incremental decision-making for exactly the embodied setting the paper targets.\n\nThe review is not systematic, so 'very little research' is supported by the authors' selection rather than a reproducible search. That is a real limitation but not fatal—they carefully say 'very little' rather than 'none'.\n\nWho should read it: researchers working on LLMs for HRI, and anyone building spoken dialogue systems that need to handle barge-in, self-correction, or early action. They will get a solid map of the incremental dialogue landscape and a useful checklist. I would send this to review, with a request to fix the DM boundary and address the timing literature. The paper is worth engaging, and the fix is not deep surgery.","headline":"A genuinely useful review whose central claim is more brittle than it looks: the 'very little incremental DM' finding depends on where you draw the DM boundary, and the paper draws it on both sides.","tokens_in":27011,"tokens_out":2836,"would_cite":true,"duration_ms":25896,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T50","68T40"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that the near-absence of incremental dialogue management is the bottleneck for natural human-robot conversation, and that large language models cannot fill the gap because their processing is monotonic.","keywords":["spoken dialogue systems","incremental processing","dialogue management","human-robot interaction","large language models","non-monotonicity","turn-taking","dialogue state tracking"],"falsifier":"A systematic bibliographic search that counts systems with explicit, timing-aware decision-making modules operating at word-level increments, regardless of whether the authors label them dialogue managers, would settle it: if such systems are numerous, the scarcity claim is false. A smaller behavioral test is to build two robots, one whose timing decisions are made by the dialogue manager and one where timing is fixed by ASR endpointing, and measure perceived naturalness and task completion; if no advantage appears, the paper's central motivation fails.","tokens_in":25964,"feed_emoji":"🤖","tokens_out":6935,"duration_ms":58727,"temperature":0.7,"pith_summary":"Spoken dialogue systems have been made incremental piece by piece: speech recognition emits words as they are spoken, language understanding updates its interpretation word by word, and language generation can add, revoke, and purge output. This paper asks where the corresponding work stands for the dialogue manager, the component that decides what the system does next, and reports that almost none exists. The authors argue this gap is now critical, because large language models are not fully incremental: their processing is monotonic and oriented toward complete sentences, so they cannot revise an interpretation or an utterance when newer observations contradict it. The paper's contribution is a set of requirements for practical incremental dialogue management, centered on owning when to act, acting on incomplete information, making fast small decisions concurrently, and moving LLMs from restart-incremental to update-incremental processing.","feed_headline":"Why talking robots stall: the dialogue manager is not incremental","feed_subtitle":"Speech modules already work word-by-word; the dialogue decision-maker barely does, and LLMs can't revise yet.","key_machinery":"The load-bearing distinction is restart-incremental versus update-incremental processing. A restart-incremental model re-reads the growing prefix on each new input and recomputes; an update-incremental model maintains a state that is revised with each word. The paper also leans on the incremental unit framework, in which each piece of information produced by a module is a node in a network that can be added, revoked, or committed. Revocation is the mechanism that makes non-monotonic dialogue possible, and it is exactly what the dialogue manager must absorb: when a recognized hypothesis such as gray is replaced by green, the manager's state and action plans need to change. This is the lens through which the authors judge LLMs, whose autoregressive output is incremental but whose committed tokens cannot be taken back, and through which they frame their requirements.","core_discovery":"At the center is a scarcity result. Surveying the literature on interactive systems, the authors find ample incremental work in automatic speech recognition, natural language understanding, natural language generation, and turn-taking, but state plainly that there is very little research on incremental dialogue management. They define dialogue management as dialogue state tracking plus action and response selection, and their desiderata explicitly assign it the additional responsibility of timing: deciding when to act, not just what to do. The papers that do exist, including an incremental information-state manager, the DIUM manager that produces self-corrections, a time-board manager, and policy-learning studies in fast-paced games, are described as promising but largely abandoned. The paper's positive claim is that this neglected component becomes the bottleneck when robots interact with people, because a robot must be able to act before the user finishes speaking, revise a plan when a recognized word is revoked, and signal understanding through early action.","pith_inferences":["The reported scarcity may partly reflect where the authors draw the line: if timing and revision are distributed across ASR endpointing, NLG output buffers, and robot-level replanners, then the literature contains more incremental decision-making than a narrow dialogue-manager search will find.","The desiderata imply an evaluation metric the community does not yet have: a cost-sensitive score that rewards acting early and correcting course, rather than only final task success or final response accuracy.","The robot replanning literature, which already handles revoking and replacing action plans under new observations, is a natural source of mechanisms for the revision half of incremental dialogue management.","If full-duplex LLM dialogue agents continue to improve, the paper's requirements suggest a concrete test: does the model decide when to act from a maintained, revisable state, or does it simply wait for a complete utterance?"],"forward_implications":["Incremental dialogue management would let a robot begin moving toward an object while the user is still finishing the request, using the motion itself as a backchannel that signals understanding.","A proper incremental dialogue manager should replace reliance on ASR endpointing: pauses would become one signal among many, with the manager deciding when silence or content is enough to act.","Evaluating dialogue managers incrementally requires new data: datasets annotated at the word level for when a decision should be made, not only for what the final decision is.","LLM-based dialogue managers must be made update-incremental, or paired with an external mechanism that can retract and repair output, rather than restarting on every growing prefix."],"supporting_citations":[{"why":"Defines the incremental unit framework with add, revoke, and commit operations that the paper uses to characterize non-monotonic processing.","marker":"Schlangen and Skantze (2011)"},{"why":"Supplies the four canonical dialogue management objectives, including deciding what to express next and when to express it.","marker":"Traum and Larsson (2003)"},{"why":"The first incremental information-state dialogue manager, evaluated against a non-incremental version.","marker":"Buß et al. (2010)"},{"why":"DIUM, the line of incremental dialogue management with explicit self-corrections, which the paper says was not pursued further.","marker":"Buß and Schlangen (2011)"},{"why":"Shows an incremental interaction manager coupling incremental ASR with a POMDP-based dialogue manager.","marker":"Selfridge and Arizmendi (2012)"},{"why":"Demonstrates an incremental dialogue policy learned with reinforcement learning for timing decisions in a fast-paced game.","marker":"Manuvinakurike et al. (2017)"},{"why":"Introduces the Time Board approach to dialogue management that coordinates past and future events.","marker":"Yaghoubzadeh et al. (2015)"},{"why":"Frames the division of decision-making responsibilities in conversational robots, which motivates the focus on the dialogue manager.","marker":"Lison and Kennington (2023)"}],"fun_headline_variants":["Incremental dialogue managers are rare, leaving robots unable to revise","Robots can't act mid-sentence—dialogue managers lag behind speech","The bottleneck for talking robots: dialogue management isn't incremental","Speech is word-by-word, but dialogue decisions can't keep up","Why robots stall: missing incremental dialogue managers, not speech modules"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scarcity finding rests on a definitional premise: dialogue management is defined as state tracking plus action selection, and timing decisions are assigned to the dialogue manager; if timing and revision are instead handled by adjacent components such as ASR endpointing, NLG output buffers, or robot-level replanners, the claimed gap could be an artifact of that division of labor.","fun_headline_variants_meta":{"raw":{"variants":["Incremental dialogue managers are rare, leaving robots unable to revise","Robots can't act mid-sentence—dialogue managers lag behind speech","The bottleneck for talking robots: dialogue management isn't incremental","Speech is word-by-word, but dialogue decisions can't keep up","Why robots stall: missing incremental dialogue managers, not speech modules"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1346,"prompt_tokens":891,"completion_tokens":455,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":365}},"tokens_in":507,"tokens_out":455,"duration_ms":5086,"temperature":1.0,"reasoning_tokens":365,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:39:42.958718+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A systematic bibliographic search that counts systems with explicit, timing-aware decision-making modules operating at word-level increments, regardless of whether the authors label them dialogue managers, would settle it: if such systems are numerous, the scarcity claim is false. A smaller behavioral test is to build two robots, one whose timing decisions are made by the dialogue manager and one where timing is fixed by ASR endpointing, and measure perceived naturalness and task completion; if no advantage appears, the paper's central motivation fails.","supporting_citations":[],"review_version":1}