{"id":"c9dc2421-f94f-402f-8e15-503b1e7a6446","arxiv_id":"2412.01262","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"ReAct-prompted GPT-3.5 and GPT-4 underperform classical task-oriented dialogue systems on task success, but humans rate them as more satisfying despite lower success.","lead":"This preprint applies ReAct prompting, a chain-of-thought plus tool-use strategy, to large language models for task-oriented dialogue in the MultiWOZ benchmark. It finds that ReAct models finish tasks far less often than classical baselines, yet human users rate them as more satisfying to talk to.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The human-preference claim rests on 95 dialogues from 20 unblinded in-house volunteers with no inferential statistics; the abstract's central contrast may be sampling noise.","rationale":"The simulation evaluation is the paper's strongest evidence: 1000 dialogues, a fixed seed, standard ConvLab metrics, and a clear internal ordering (GPT-4 above GPT-3.5; both below HDC/RL baselines). The cost analysis and qualitative failure analysis are useful and honestly reported. The load-bearing weakness is exactly where the reader placed it: the human evaluation. The abstract's positive twist, that humans prefer ReAct-LLM despite lower task success, rests on a single underpowered, unblinded, in-house study with no inferential statistics. The 20 volunteers are not described, the assignment mechanism is under-specified, and repeated dialogues per user violate independence. Because the satisfaction scale is ordinal and converted to percentages, reporting only means hides the distribution and makes the result hard to interpret. I would not call the claim false; it may well be true. But the current evidence does not establish it beyond sampling noise. The proposed causal explanation (fluency, politeness, confidence) is speculative, as the paper itself hedges with 'most likely.' This warrants the same CONDITIONAL verdict the reader gave: the central human-preference claim should be accepted only with raw-data reanalysis or a larger, pre-registered human study. I therefore agree with the reader's weakest assumption and do not change the verdict.","tokens_in":12025,"tokens_out":5172,"duration_ms":49080,"concrete_test":"Obtain the raw per-dialogue and per-user satisfaction ratings from the authors and rerun the comparison with a cluster bootstrap that resamples users (not dialogues), computing a 95% confidence interval for the mean satisfaction difference (ReAct-LLM minus HDC) and a permutation test under the null of no system effect. If the confidence interval includes zero or the permutation p-value exceeds 0.05, the headline human-preference claim is not statistically established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that 'humans report higher subjective satisfaction with ReAct-LLM despite its lower success rate' (Abstract; Section 5.2, Table 3) depends entirely on the human evaluation. With 20 in-house volunteers producing 95 dialogues per system, the paper reports only raw percentage averages: satisfaction 54.10% vs 65.47%, success 60.00% vs 50.52%. No confidence intervals, significance tests, effect sizes, or inter-annotator agreement are given. Because each volunteer contributed multiple dialogues and systems were 'randomly chosen' per session, the 190 observations are not independent; the effective sample size for the satisfaction comparison is closer to 20 users than to 190 dialogues. An 11.37-point mean difference on a 1-5 scale converted to percentages (about 0.57/5) could easily arise from a few outlier users or from order/learning effects, especially since 'users could start a conversation as many times as they wish' (Section 5.2). Moreover, the proposed explanation that ReAct-LLM is preferred because it is 'self-confident and generates well structured, polite, fluent and natural sentences' is an untested post-hoc attribution; no rating dimension beyond overall satisfaction was collected to support it. If this preference claim is unsupported, the abstract's contrast between simulation and human evaluation loses its main positive result, leaving only the negative simulation finding and a qualitative error analysis. The simulation result itself is less at risk: 1000 fixed-seed dialogues with replicable ConvLab metrics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a ReAct-based LLM system for task-oriented dialogue (TOD) using GPT-3.5 and GPT-4, with explicit tool calls for domain selection, slot extraction, database queries, and booking reference generation, and evaluates it on MultiWOZ using the ConvLab-3 simulator. In 1000 simulated dialogues, ReAct-LLM reaches success rates of 28.2% (GPT-3.5) and 43.6% (GPT-4), far below the HDC (83.8%) and ACGOS (81.7%) baselines reported in Table 1. In a human evaluation with 95 dialogues per system from 20 in-house volunteers, HDC still achieves a higher success rate (60.0% vs 50.5%), but users report higher overall satisfaction with ReAct-LLM (65.5% vs 54.1% after converting 1-5 ratings to percentages). The paper also provides a qualitative error analysis of reasoning traces, hallucinated slots, and failures of an LLM-based user simulator, plus an API-cost analysis.","tokens_in":12309,"tokens_out":6017,"duration_ms":49193,"significance":"If the human-preference finding were rigorously established, the paper would make a noteworthy contribution: users may prefer fluent, confident LLM-generated responses even when task success is lower, and simulation-based comparisons may systematically undervalue LLM-based TOD systems. The simulated evaluation is reproducible (fixed seed, standard ConvLab-3 metrics, multiple external baselines) and the qualitative analysis of failure modes is informative. The cost analysis is also practically useful. However, the central positive claim about human satisfaction rests on a small, unblinded, in-house sample with no inferential statistics, so the paper's headline result is not yet supported at the level expected by a journal. The paper is a valuable empirical study in need of a more rigorous human evaluation.","major_comments":[{"comment":"The claim that 'users are more satisfied with ReAct-LLM than with HDC' is not statistically supported. The data come from 95 dialogues per system collected from 20 users, but the observations are not independent because each user contributed multiple dialogues, and no confidence intervals, significance tests, effect sizes, or inter-annotator agreement are reported. The satisfaction difference is 11.37 percentage points (54.10% vs 65.47%, i.e., about 0.57 on a 1-5 scale) and could plausibly be driven by a few outlier users or by practice/fatigue effects, especially since users could start a conversation as many times as they wished. The authors should provide a per-participant analysis (e.g., mean satisfaction per user and per system), a significance test that accounts for repeated measures (e.g., Wilcoxon signed-rank test or a mixed-effects model with a random intercept for participant), confidence intervals, and robustness checks excluding outliers. Without these, the abstract's main positive result is not established.","section":"Section 5.2, Table 3"},{"comment":"The explanation that ReAct-LLM is preferred 'because ReAct-LLM is self-confident and generates well structured, polite, fluent and natural sentences' is a post-hoc attribution with no supporting rating data. No rating dimension beyond overall satisfaction was collected, so the paper cannot distinguish whether fluency, politeness, naturalness, confidence, or some other factor drives the preference. The authors should either collect per-dimension ratings (e.g., fluency, politeness, naturalness, confidence) or present this explanation as a hypothesis for future work rather than as a finding of the current study.","section":"Section 5.2"},{"comment":"The simulation comparison is potentially confounded by the user simulator's BERT-NLU errors when parsing ReAct-LLM's free-form responses. The paper acknowledges this in the conclusion ('difficulty in understanding the system requests by the user simulator, due to BERT-NLU errors') but does not quantify how many simulated failures are caused by NLU misparsing versus ReAct reasoning or response-generation errors. Since the abstract's first claim ('severely underperform ... on success rate in simulation') is based on Table 1, the authors should provide a breakdown of failure causes on a manually annotated sample of the 1000 simulated dialogues, or otherwise estimate the effect of NLU errors on the reported success rates.","section":"Section 5.1 and Section 6"},{"comment":"The human evaluation protocol is under-specified, making the result difficult to interpret or reproduce. The paper does not report how many dialogues each user contributed, whether the same goals were used for both systems, whether the users were blind to the system identity, how repeated conversation attempts were handled in the success computation, what instructions were given for assigning the satisfaction rating, or how the 95 dialogues per system were distributed across the 20 users. These details should be provided in a revised version, as they bear directly on the validity of the satisfaction comparison.","section":"Sections 4.2 and 5.2"}],"minor_comments":[{"comment":"The header of Table 3 appears garbled: 'Avg Satisfaction Success Turn Rate (%) Rate (%)' should be relabeled as separate columns, e.g., 'Satisfaction Rate (%)', 'Success Rate (%)', and 'Avg Turns'.","section":"Table 3"},{"comment":"The paper states that a fixed random seed was used in the goal generator for reproducibility, but does not report the seed value; the seed should be stated explicitly.","section":"Section 4.2"},{"comment":"There are inconsistent spellings of 'ReAct' (e.g., 'React-LLMs' in Section 5.1) and 'ConvLab-3' (written as 'CONVLAB 3' in Section 4.2); please standardize these terms.","section":"Throughout"},{"comment":"In the ReAct example, the input to db_query is shown as '{domai: hotel, state: ...}'; this appears to be a typo for '{domain: hotel, ...}'.","section":"Figure 3"},{"comment":"The sentence 'the performance of GPT-4 is superior to GPT-3' should refer to GPT-3.5, not GPT-3, to avoid confusion with the actual models used.","section":"Appendix D"},{"comment":"The prompt says 'when you do only specify the tool name as the Action' but the accompanying example includes an 'Input' field for each action; please clarify the intended format for tool arguments.","section":"Figure 2"},{"comment":"The human evaluation section does not mention whether ethical approval or informed consent was obtained for the in-house volunteers; this information should be added if applicable.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is clearly written, and the simulation setup is solid and reproducible. My main concern is that the human-preference claim, which is the paper's most interesting contribution, is currently supported only by raw averages from a small, unblinded sample with no inferential statistics. If the authors can add a proper per-participant statistical analysis, strengthen the protocol reporting, and temper the post-hoc explanation, the paper would likely be acceptable. The fit with the journal is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this paper. First, it does something genuinely new: it applies ReAct prompting to full MultiWOZ dialogues in ConvLab3 and runs both simulated and human evaluations. The simulation result is sobering and credible: ReAct-GPT-4 reaches 43.6% success versus 83.8% for a BERT-NLU+HDC+template baseline, and the gap is even larger against semantic-level baselines. That negative result is worth knowing. Second, the paper's more interesting claim—that humans rate ReAct-LLM higher in satisfaction despite lower task success—rests on a very thin human study. With 20 in-house volunteers, 95 dialogues per system, no significance tests, no confidence intervals, and no inter-annotator agreement, the 11-point satisfaction gap could easily be noise. The explanation offered (confidence, politeness, fluency) is post-hoc, with no rating dimension measured to support it.\n\nWhat the paper does well: the experimental setup is clearly described, the seed is fixed, metrics are standard, and the cost analysis is a nice addition. The qualitative error analysis (hallucinated slots, not sticking to tools, not asking clarifying questions) is useful for anyone building LLM agents for TOD. The authors also honestly report that an LLM user simulator failed and why, which is informative.\n\nWhere it's soft: the human evaluation is the load-bearing wall for the abstract's central contrast, and it doesn't hold up statistically. The authors need either a larger, independent sample with proper inference, or a more cautious framing that presents the preference result as an observation rather than a finding. Also, the comparison would be fairer if the human study used GPT-4 as well, since the simulation shows GPT-4 is the better model; the human study only tests GPT-3.5. Minor point: the domain-specific prompt experiment is underpowered (100 dialogues) and they note no effect, which is fine but not very informative.\n\nOverall, this is a solid within-subfield paper with a clear negative simulation result and an intriguing but under-powered human finding. It deserves peer review—the negative result alone is publishable—but the human claim needs to be either made rigorous or clearly demoted to an anecdote.\n\nRecommendation: send to review, but expect substantial revision on the human evaluation section.","headline":"Solid simulation study showing ReAct underperforms in TOD, but the human preference claim is too underpowered to support the abstract's main contrast.","tokens_in":12851,"tokens_out":2349,"would_cite":true,"duration_ms":20964,"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":"ReAct-prompted LLMs lose on task success but win on user satisfaction in task-oriented dialogue, and the paper attributes the reversal to fluent, confident phrasing.","keywords":["task-oriented dialogue","ReAct prompting","large language models","LLM agents","user satisfaction","human evaluation","MultiWOZ","dialogue state tracking"],"falsifier":"A targeted refutation would be a larger, preregistered human study: with at least 50 users and 200 or more dialogues per system, compute a paired significance test on the satisfaction difference (65.47% vs 54.10%). If the difference is not statistically significant, or if it flips when users are told whether their goal was actually achieved, the paper's central human-evaluation conclusion would not hold.","tokens_in":11837,"feed_emoji":"🤖","tokens_out":7776,"duration_ms":63228,"temperature":0.7,"pith_summary":"Task-oriented dialogue systems usually need hand-built pipelines or large training corpora; the paper asks whether ReAct prompting can make a general-purpose LLM do the whole job. It builds a ReAct-LLM agent on GPT-3.5 and GPT-4 with four database tools, and compares it with classical dialogue-management baselines on 1,000 simulated dialogues. In simulation the LLM agents complete far fewer user goals than the baselines. In a human study with 20 users and 95 dialogues per system the success gap narrows, and users rate the ReAct system more satisfying even though it succeeds less often. The paper concludes that fluent, confident phrasing can carry higher satisfaction than task success.","feed_headline":"Dialogue bot that fails more tasks still wins user satisfaction","feed_subtitle":"In a human study, ReAct-prompted GPT-3.5 scored 65% satisfaction versus 54% for a higher-success baseline.","key_machinery":"The mechanism is ReAct prompting, a few-shot prompting loop in which the LLM alternates Thought (reasoning about what to do), Action (choosing a tool), and Observation (reading the tool's result) before producing a final answer. The agent is given four tools that expose the MultiWOZ database: list_domains, list_slots, db_query, and get_booking_reference. The intended role of this loop is to substitute for the classical pipeline steps of domain selection, slot detection, belief-state tracking, database query, and response generation, with a single worked example in the prompt as the only training signal.","core_discovery":"The paper's central claim is that ReAct prompting alone—no fine-tuning, no bespoke dialogue state tracker—can make an LLM act as a task-oriented dialogue agent, but with a stark evaluation split. On the ConvLab-3 simulation benchmark the ReAct-LLM reaches only 28.2% success with GPT-3.5 and 43.6% with GPT-4, versus 83.8% for the BERT-NLU + HDC baseline; with semantic-level interaction the baselines reach over 97%. With real users, HDC completes 60.0% of goals versus 50.5% for ReAct-LLM, yet users give ReAct-LLM a 65.5% average satisfaction rating versus 54.1% for HDC. The authors attribute this to the LLM's self-confident, well-structured, polite and fluent sentences, and identify recurring failure modes: reasoning that imitates the prompt example, inconsistent or hallucinated slots, poor belief-state accumulation, and missing clarifying questions.","pith_inferences":["The paper does not test this, but a follow-up that matched phrasing style between systems could isolate whether fluency alone explains the satisfaction gap.","The results imply that user-satisfaction ratings should not be treated as a proxy for task success in dialogue evaluation; the trends point in opposite directions.","The LLM user simulator's role-switching failure suggests that LLM-as-user evaluation is fragile, so future agent benchmarks may need stricter user-simulation protocols."],"forward_implications":["ReAct prompting by itself is not a drop-in replacement for classical TOD pipelines when the metric is task completion: simulated success rates are roughly 28–44% versus 84% for a BERT-NLU + HDC baseline.","Benchmarks built on simulated users can understate an LLM's perceived quality by real users; the paper's human data show satisfaction can be high even when task success is low.","The per-token cost of GPT-4 is about 36 times that of GPT-3.5 in this setup, and the gain in success rate is not enough to justify it in the authors' assessment.","LLM-based dialogue agents need additional control mechanisms to keep reasoning traces and belief states valid; the qualitative analysis shows hallucinations, invalid slots, and irrelevant actions as concrete failure modes."],"supporting_citations":[{"why":"Defines the ReAct reasoning-and-acting prompting strategy that the paper adapts to task-oriented dialogue.","marker":"(Yao et al., 2022)"},{"why":"Provides the MultiWOZ dataset, domains, and goal structure used to generate the evaluation dialogues.","marker":"(Budzianowski et al., 2018)"},{"why":"Supplies the ConvLab-3 toolkit, the agenda-based simulator, and the success, book, inform, and complete metrics.","marker":"(Zhu et al., 2022)"},{"why":"Introduces the agenda-based user simulation method that the ConvLab simulator runs in the simulated evaluation.","marker":"(Schatzmann et al., 2007)"},{"why":"Provides the ACGOS structured-RL baseline and imitation-learning policy that the paper compares against ReAct-LLM.","marker":"(Cordier et al., 2022)"},{"why":"The GPT-4 model used in the ReAct-LLM agent, and the source of the model's reported behavior in qualitative analysis.","marker":"(OpenAI, 2023)"},{"why":"The instruction-following language model behind GPT-3.5, the cheaper ReAct-LLM variant used in both evaluations.","marker":"(Ouyang et al., 2022)"}],"fun_headline_variants":["LLM underperforms tasks yet outranks in user satisfaction","Fails more tasks, still preferred by users","Simulation flop, human win: ReAct's split verdict","ReAct prompts: worse success, better user ratings","Task success drops, satisfaction rises with ReAct"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the satisfaction scores of 20 in-house volunteers, collected over 95 dialogues per system without significance tests or confidence intervals, are representative enough to show that real users prefer ReAct-LLM over the HDC baseline despite its lower task success.","fun_headline_variants_meta":{"raw":{"variants":["LLM underperforms tasks yet outranks in user satisfaction","Fails more tasks, still preferred by users","Simulation flop, human win: ReAct's split verdict","ReAct prompts: worse success, better user ratings","Task success drops, satisfaction rises with ReAct"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000184,"raw_usage":{"total_tokens":1298,"prompt_tokens":904,"completion_tokens":394,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":314}},"tokens_in":520,"tokens_out":394,"duration_ms":3974,"temperature":1.0,"reasoning_tokens":314,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:30:31.146657+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A targeted refutation would be a larger, preregistered human study: with at least 50 users and 200 or more dialogues per system, compute a paired significance test on the satisfaction difference (65.47% vs 54.10%). If the difference is not statistically significant, or if it flips when users are told whether their goal was actually achieved, the paper's central human-evaluation conclusion would not hold.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the agenda-based user simulation method that the ConvLab simulator runs in the simulated evaluation."}],"review_version":1}