{"id":"d921e58c-5ddb-4808-aed7-40108df680f9","arxiv_id":"2611.25773","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Duopra trains a full-duplex speech model to initiate and execute tool calls, raising Pass@1 from 4.7% to 44.7% on FDB v3 by overcoming the frontend initiation bottleneck.","lead":"Duopra is a full-duplex voice interaction model that learns to call tools directly or delegate tasks to a backend, raising task completion from 4.7% to 44.7% on the Full-Duplex-Bench v3 benchmark. The paper shows that the main bottleneck in frontend-backend voice agents is initiating a faithful request, not backend execution.","discovery_kind":"new_method","skeptic_critique":{"audit":{"quote":null,"status":"unresolved_evidence","question":"Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.","consequence":null,"objection_id":"O1","published_as":"question"},"model":"deepseek/deepseek-v4-flash","headline":"Not confirmed. Pith could not check this concern against the text of the paper it was given; it is listed as an open question, not a finding.","rationale":"","tokens_in":48193,"tokens_out":502,"duration_ms":9739,"concrete_test":"","verdict_should_be":null,"load_bearing_attack":"Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.","agreement_with_reader":null},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper introduces Duopra, a full-duplex interaction model that learns task execution alongside conversational control. It treats the backend as a tool and uses a three-stage training process (protocol SFT, full-duplex RL, projector-only repair). On FDB v3, Duopra raises Pass@1 from 4.7% to 44.7% with the same Qwen3-8B backend, and to 49.3% with a DeepSeek backend. The paper analyzes that the main bottleneck in frontend-backend agents is initiation (unsent requests), not backend execution.","tokens_in":48521,"tokens_out":2113,"duration_ms":27503,"significance":"Not published. Pith's referee wrote this assessment before its objections were checked against the paper, and 5 of its 5 major objections did not survive that check.","major_comments":[],"minor_comments":[{"quote":null,"comment":"Figure 1 is described but not visible in the supplied text. The caption is informative, but the figure itself could not be inspected.","section":"§1, Figure 1","consequence":null,"audit_status":null,"objection_id":null,"quote_matched":null,"required_revision":null,"authors_acknowledge":null,"authors_acknowledge_p":null},{"quote":"At unit~$t$, the frontend reads $x_t$ and the observations $e_t$ injected by then, which hold the schemas of the tools it can call and the results of its earlier calls","comment":"The policy equation uses notation $x_{\\le t}$, $e_{\\le t}$, $y_{<t}$. The definition of $e_t$ (observations) could be clarified earlier; it is introduced in §4.1 but used in §3.","section":"§3, Eq. (1)","consequence":null,"audit_status":null,"objection_id":null,"quote_matched":true,"required_revision":null,"authors_acknowledge":null,"authors_acknowledge_p":null},{"quote":"We write $\\bar\\rho_{ij}$ for this ratio clipped to $[1-\\epsilon,1+\\epsilon]$, and $K_{ij}$ for the per-token KL estimator against Duopra-SFT.","comment":"The KL estimator $K_{ij}=v_{ij}-\\log v_{ij}-1$ is defined in §A.2 but referenced in §4.3. The definition should be given earlier or the reference should be clearer.","section":"§4.3, Eq. (3)","consequence":null,"audit_status":null,"objection_id":null,"quote_matched":true,"required_revision":null,"authors_acknowledge":null,"authors_acknowledge_p":null},{"quote":"$^{m}$Tool-call success, not Pass@1. $^{h}$Tool-selection accuracy, not F1. $^{x}$On VoiceChat (49.0 on Qwen-Realtime), with exact-string argument matching. $^{o}$Offline inference with tool definitions in the prompt.","comment":"The table uses superscripts (m, h, x, o) for footnotes, but the footnotes are at the bottom of the table and may be missed. Consider using a more prominent notation.","section":"§5.2, Table 1","consequence":null,"audit_status":null,"objection_id":null,"quote_matched":true,"required_revision":null,"authors_acknowledge":null,"authors_acknowledge_p":null},{"quote":"The scorer validates the output schema, verifies that quotes occur in the public text, and checks that each cited observation supports the specified goal and precedes the statement.","comment":"The reward-judge instruction in Table 9 is very long and complex. The paper should summarize the key validation steps that ensure the judge's reliability.","section":"§A.5, Table 9","consequence":null,"audit_status":null,"objection_id":null,"quote_matched":true,"required_revision":null,"authors_acknowledge":null,"authors_acknowledge_p":null},{"quote":"Successes by required call count use 198 one-call, 54 two-call and 48 three-call sessions.","comment":"The table reports 'Passes by calls' for 1, 2, 3 required calls. The total number of sessions for each call count is 198, 54, 48 respectively (from 300 sessions total). This should be stated in the caption or a note.","section":"§A.8, Table 10","consequence":null,"audit_status":null,"objection_id":null,"quote_matched":true,"required_revision":null,"authors_acknowledge":null,"authors_acknowledge_p":null},{"quote":"No backchannels are produced by any of the three post-trained checkpoints on the 55 FDB v1.0 backchannel items; the reported frequency and TOR are zero, and JSD is 1.000.","comment":"The backchannel TOR and frequency are reported as 0.0 for all post-trained checkpoints. The paper should clarify whether this is because the model never produces backchannels or because the evaluation metric does not detect them.","section":"§A.9, Table 11","consequence":null,"audit_status":null,"objection_id":null,"quote_matched":true,"required_revision":null,"authors_acknowledge":null,"authors_acknowledge_p":null},{"quote":"\\paragraph{Comparability of published results.} We include published results for commercial systems and several open-source baselines as reference points.","comment":"The limitations section is well-written and acknowledges key issues (comparability, model coverage, fixed user input). However, it does not mention the potential circularity of using the same model family for reward and evaluation.","section":"§6, Limitations","consequence":null,"audit_status":null,"objection_id":null,"quote_matched":true,"required_revision":null,"authors_acknowledge":null,"authors_acknowledge_p":null}],"open_questions":[{"origin":"referee","reason":"source_unmatched","section":"§5.2, Table 1; §A.6, Table 9","question":"Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.","objection_id":"O2"},{"origin":"referee","reason":"source_unmatched","section":"§5.2, Table 1; §A.6, Table 9","question":"Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.","objection_id":"O3"},{"origin":"referee","reason":"source_unmatched","section":"§5.3, Table 2; §A.8, Table 10","question":"Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.","objection_id":"O4"},{"origin":"referee","reason":"source_unmatched","section":"§4.3, Eq. (3); §A.3, Table 5","question":"Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.","objection_id":"O5"},{"origin":"referee","reason":"source_unmatched","section":"§5.4, Table 3; §A.9, Table 11","question":"Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.","objection_id":"O6"}],"recommendation":"uncertain"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"This paper makes a concrete empirical contribution. The central finding is that in frontend-backend voice agents, the bottleneck is not backend execution but frontend initiation: the model fails to send a request at all. Duopra addresses this by training a single full-duplex policy to emit tool calls and backend delegations through a unified action space, using a three-stage recipe of protocol SFT, execution-based RL, and projector-only talker repair. The result is a large and clean gain on FDB v3 (4.7% to 44.7% Pass@1 with the same Qwen3-8B backend), and the decomposition into Call and Pass|call convincingly shows where the improvement comes from.\n\nThe experimental design is thorough: three decoding seeds, paired bootstrap intervals, multiple baselines, ablations, and analysis of failure modes. The paper is honest about its limitations, including the use of synthetic speech, prerecorded requests, and a single base model. The loss of spontaneous backchanneling is reported rather than hidden.\n\nThe main soft spot is the reliance on the same LLM (Qwen3.8-27B) as both the reward judge during training and the evaluation judge for argument and response scores. The paper uses a detailed structured judgment protocol and validates the judge's output, but the potential circularity is not addressed. If the judge systematically favors the model's outputs, the reported scores could be inflated. This is a real concern, though the Pass@1 metric (which depends on exact call matching, not the judge) provides a partial check. A second concern is that the projector repair evaluation uses selection-set evidence, so the reported CER/WER reductions are not independent generalization estimates. The paper acknowledges this.\n\nWho this is for: researchers working on full-duplex voice agents, tool use in spoken dialogue, and post-training for speech models. The paper deserves a serious referee. The authors should address the judge circularity concern, ideally by evaluating on a subset with a different judge or by providing evidence that the judge's errors are not systematically biased toward Duopra.","headline":"Duopra shows that a full-duplex voice frontend can learn to initiate tool calls and backend requests, raising task completion from 4.7% to 44.7% on FDB v3, and the paper's analysis cleanly separates the initiation bottleneck from execution quality.","tokens_in":49052,"tokens_out":513,"duration_ms":8833},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"The paper argues that task success in full-duplex voice agents is limited first by initiation, not execution, and introduces Duopra to learn both.","keywords":["full-duplex interaction","voice agents","tool use","task agency","reinforcement learning","speech language model","frontend-backend architecture","initiation bottleneck"],"falsifier":"If an independent evaluation using a different judge (e.g., GPT-4o as in the original FDB v3 benchmark) or human annotation showed that Duopra's Pass@1 gain over RV-Audio + Harness is substantially smaller (e.g., less than 20 percentage points), the central claim that initiation is the primary bottleneck and that Duopra recovers unserved requests would be weakened.","tokens_in":50077,"feed_emoji":"🎙️","tokens_out":1631,"duration_ms":19512,"temperature":0.2,"pith_summary":"The paper investigates why frontend–backend full-duplex voice agents fail to complete tasks. It finds that the bottleneck is not the backend's ability to execute tool calls, but the frontend's failure to initiate a faithful request in the first place. The authors introduce Duopra, a model that learns to make direct tool calls and to delegate through a unified action interface, trained with a three-stage procedure combining supervised fine-tuning, reinforcement learning with time-anchored rewards, and projector-only repair to preserve speech quality. On the Full-Duplex-Bench v3 benchmark, Duopra raises Pass@1 from 4.7% (RV-Audio + Harness) to 44.7% with the same Qwen3-8B backend, and to 49.3% with a DeepSeek backend, while also improving conversational takeover rates. The central claim is that task agency—deciding when and how to act—can be learned alongside conversational agency in a single causal policy.","feed_headline":"Authors report that task success in voice agents is limited by initiation, not execution","feed_subtitle":"Duopra learns to initiate tool calls and delegations, raising Pass@1 from 4.7% to 44.7% on FDB v3","key_machinery":"The central mechanism is the unified action interface that treats both direct tool calls and backend delegation as instances of the same \texttt{call} operation, with the backend represented as a tool \texttt{query(task)} accepting a self-contained natural-language argument. This allows the same causal streaming policy $\\pi_\\theta(y_t \\mid x_{\\le t}, e_{\\le t}, y_{<t})$ to decide turn states, speech tokens, and private actions jointly. The three-stage training procedure is the key machinery: Stage 1 (protocol SFT) teaches schema-based calling and clarification; Stage 2 (full-duplex RL with Dr. GRPO) rewards timely, complete, and grounded execution using time-anchored deadlines and outcome tiers; Stage 3 (projector-only repair) realigns talker conditioning after thinker updates to mitigate speech degradation.","core_discovery":"The paper establishes that in frontend–backend full-duplex voice agents, task success is limited first by initiation, not execution. The frontend must decide whether and what to request; unsent requests cannot be executed. The authors demonstrate that a full-duplex interaction model can learn to initiate faithful tool calls and backend delegations through a unified action interface, recovering tasks that a baseline frontend never attempts. Duopra's three-stage training—protocol SFT, full-duplex RL with time-anchored deadlines and outcome tiers, and projector-only talker repair—enables the model to achieve 44.7% Pass@1 on FDB v3 with the same Qwen3-8B backend that the baseline uses, compared to 4.7% for the baseline. The gain comes almost entirely from sessions the baseline never executed, confirming that initiation is the primary bottleneck.","headline_check":{"rule":"report_a_claim_v1","sent_back":false,"unattributed":{}},"pith_inferences":["The paper's decomposition of Pass@1 into Call × Pass|call provides a diagnostic tool that could be applied to other frontend–backend systems to identify whether failures stem from initiation or execution.","The loss of spontaneous backchanneling after task training suggests a tension between task-oriented and purely conversational behaviors that may require explicit multi-objective optimization or data mixing to resolve.","The projector-only repair technique (aligning talker conditioning after thinker updates) could be adopted by other speech models that fine-tune a language model component and need to preserve speech quality without retraining the full speech generator.","The paper's information-sufficiency criterion (a call is justified only if each required argument has a source) formalizes a key aspect of task agency that could be used as a training signal or evaluation metric in other voice agent systems."],"forward_implications":["If Duopra's approach generalizes, full-duplex voice agents can be trained to handle multi-step tool use, corrections, and conditional dependencies without relying on a separate dispatcher or backend for initiation decisions.","The finding that initiation is the primary bottleneck implies that improving backend execution alone will yield diminishing returns; effort should focus on frontend task agency.","The three-stage training recipe (SFT + RL + projector repair) provides a template for post-training other full-duplex models to acquire task agency while retaining conversational control.","The unified action interface (direct calls and delegation through the same \texttt{call} operation) simplifies the architecture and allows the model to choose the best route without a separate routing module."],"supporting_citations":[{"why":"Supplies the FDB v3 benchmark used for evaluation, including the 100 disfluent recordings, 12 simulated APIs, and the Pass@1 metric.","marker":"lin2026fdbv3"},{"why":"Provides the base model RV-Audio and the frontend–backend architecture (RV-Audio + Harness) that Duopra builds on and compares against.","marker":"venus2026realtimevenus"},{"why":"Supplies the FDB v1.0 benchmark used to evaluate conversational takeover rates (turn-taking, pauses, interruptions).","marker":"lin2025fullduplexbench"},{"why":"Provides the Dr. GRPO reinforcement learning algorithm used in Stage 2 of Duopra's training.","marker":"liu2025drgrpo"},{"why":"Supplies the LoRA adapter method used for efficient fine-tuning of the thinker in Stage 2.","marker":"hu2022lora"},{"why":"Supplies the L2-SP regularization term used in Stage 3's projector repair loss.","marker":"li2018explicit"},{"why":"Provides a comparable frontend–backend baseline system and analysis of delegation in full-duplex speech models.","marker":"hu2026frontendbackend"},{"why":"Provides another frontend–backend baseline (DuplexOmni) evaluated in the same streaming pipeline.","marker":"huang2026duplexomni"}],"cache_read_input_tokens":0,"weakest_assumption_plain":"The reward judge (Qwen3.8-27B) reliably evaluates argument correctness, response grounding, and fabrication in the streaming setting, and the paper relies on this judge for both training rewards and evaluation metrics.","cache_creation_input_tokens":0},"created_at":"2026-11-20T09:41:20.217962+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium+deepseek/deepseek-v4-flash+p.d60f54226f00","input_kind":"tex","input_chars":153496,"input_total_chars":153496,"passes_missing":[],"published_standing":"established_narrower","published_confidence":"moderate","objection_audit":{"version":"pith-objection-audit@1","counts":{"raised":6,"confirmed":0,"questions":6,"withdrawn":0,"moved_to_minor":0,"math_error_not_certified":0,"withdrawn_reviewer_error":0,"typesetting_minor_dropped":0,"withdrawn_already_addressed":0,"referee_assessment_unpublished":1},"open_questions":[{"origin":"skeptic","reason":"source_unmatched","section":null,"question":"Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.","objection_id":"O1"},{"origin":"referee","reason":"source_unmatched","section":"§5.2, Table 1; §A.6, Table 9","question":"Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.","objection_id":"O2"},{"origin":"referee","reason":"source_unmatched","section":"§5.2, Table 1; §A.6, Table 9","question":"Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.","objection_id":"O3"},{"origin":"referee","reason":"source_unmatched","section":"§5.3, Table 2; §A.8, Table 10","question":"Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.","objection_id":"O4"},{"origin":"referee","reason":"source_unmatched","section":"§4.3, Eq. (3); §A.3, Table 5","question":"Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.","objection_id":"O5"},{"origin":"referee","reason":"source_unmatched","section":"§5.4, Table 3; §A.9, Table 11","question":"Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.","objection_id":"O6"}],"standing":{"basis":["first_pass_verdict"],"standing":"established_narrower","confidence":"moderate"},"standing_rule":"1. No standing is published (not_assessed) when the review read only the abstract, when the referee pass or the objection check did not run, or when the first pass gave no verdict. 2. The first pass's verdict sets the start: ACCEPT is established, CONDITIONAL is established_narrower, REJECT is not_supported. 3. Each objection that survived the check against the paper's text caps the standing: one that breaks the central claim caps it at not_supported, one that narrows it at established_narrower, one that weakens it at supported_with_gaps. The standing is the lowest of the start and the caps. 4. A first-pass REJECT stands only if a surviving objection breaks the central claim. Without one, the standing is the lowest cap of the surviving objections, or not_assessed when none survived. 5. The confidence is the lowest of the first pass's confidence and the confidence of each objection that set the standing, and is no higher than moderate while a question is open or part of the paper's text was not supplied. 6. If any objection was withdrawn because Pith's reviewer misread the paper, the confidence is no higher than moderate, whatever the first pass said: a review that misread the paper once may have misread it elsewhere.","input_kind":"tex","nothing_to_audit":false,"bar":{"enabled":false}},"run_record":{"models":["deepseek/deepseek-v4-flash"],"passes":{"reader":{"model":"deepseek/deepseek-v4-flash","effort":"medium","cost_usd":0.005435,"tokens_in":52970,"tokens_out":3378,"duration_ms":41469,"input_sha256":"11cfeba32419278d0b08b13654f7e3ed69a971fbf2a7f296ef13ca5a85cfac05","prompt_sha256":"f9ead730038216984cfb56d7cb34332bdf691828296a1de4245f3d6667d4bf56","message_sha256":"af80f2f1adb75c905d290affc5a5c6458dd0330e9b55c6d31374a17617730f31","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"referee":{"model":"deepseek/deepseek-v4-flash","effort":"medium","cost_usd":0.0048,"tokens_in":48521,"tokens_out":2113,"duration_ms":27503,"input_sha256":"11cfeba32419278d0b08b13654f7e3ed69a971fbf2a7f296ef13ca5a85cfac05","prompt_sha256":"a852d99aea0ddfd900e1483f76e20777aae3afde9c016d5b8d6e5db01064c728","message_sha256":"0d33232fed9f5acace2af95482c0fd392e6bd548eb0373b0ad330a17a91fbecf","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"skeptic":{"model":"deepseek/deepseek-v4-flash","effort":"medium","cost_usd":0.004477,"tokens_in":48193,"tokens_out":502,"duration_ms":9739,"input_sha256":"11cfeba32419278d0b08b13654f7e3ed69a971fbf2a7f296ef13ca5a85cfac05","prompt_sha256":"8f88c0f61488b1e764c317998cddf35fffc33a1f20a3b1be54330d7843386fc7","message_sha256":"83fb8372c8b87ec6eb1f933e6a96687a84aa7c0055e3ae546c29f3e8d4242fff","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"circularity":{"model":"deepseek/deepseek-v4-flash","effort":"medium","cost_usd":0.004513,"tokens_in":48326,"tokens_out":634,"duration_ms":9013,"input_sha256":"11cfeba32419278d0b08b13654f7e3ed69a971fbf2a7f296ef13ca5a85cfac05","prompt_sha256":"e9946e39d67b76cc60476379bc32abdc44665b5f0c438659796547bb1041a01f","message_sha256":"0e6d2b3c2f3716e44b2593fdf0fc004c05273d3ba3a330a8cd1639314968f9c8","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"desk_editor":{"model":"deepseek/deepseek-v4-flash","effort":"medium","cost_usd":0.004557,"tokens_in":49052,"tokens_out":513,"duration_ms":8833,"input_sha256":"11cfeba32419278d0b08b13654f7e3ed69a971fbf2a7f296ef13ca5a85cfac05","prompt_sha256":"e7211f6a3d2734c1025ce7e7e6d48f78b0944d0a32d5021f81be63b27721420a","message_sha256":"ca9933d0bae0f19da5dfe61ddfc46569e58d9f16455d5ff1de742d95ddecaff5","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"objection_audit":{"model":"deepseek/deepseek-v4-flash","effort":"medium","cost_usd":0.006651000000000001,"tokens_in":99439,"tokens_out":6588,"duration_ms":60025,"input_sha256":"11cfeba32419278d0b08b13654f7e3ed69a971fbf2a7f296ef13ca5a85cfac05","prompt_sha256":"6c7a6883d5517cf9e4945d83947200a36dcd152b8b56c32baf7500d8b3841d0a","message_sha256":"d5c980e73560be8d11947653660db0157443eab70634d1f3b962f8c9708c6446","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"pith_extraction":{"model":"deepseek/deepseek-v4-flash","effort":"medium","cost_usd":0.004854,"tokens_in":50077,"tokens_out":1631,"duration_ms":19512,"temperature":0.2,"input_sha256":"11cfeba32419278d0b08b13654f7e3ed69a971fbf2a7f296ef13ca5a85cfac05","prompt_sha256":"e4f56bf568e7854d0a1d2efb6ec96f8697d5132737cdb31cb7fc64acbd923974","message_sha256":"e5d98ceec436a5076d8742cd9a543f659c7b94c39eecafae05ebaaad27ad5eb5","cache_read_input_tokens":0,"cache_creation_input_tokens":0}},"version":"pith-run-record/v1","prompt_set_sha256":"d60f54226f003b7308ab6f0f9f2951c09c726f00cc66fc05af3456f212247e93","configured_pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium"},"basis":{"kind":"full_text","source":"recorded","chars_given":153496,"chars_total":153496},"receipt":{"review_sha256":"3c9fb7fa20ecba3a5d45e71f89cc08bf99f727ab69c5c366ce7efcd23fd90795","rule":"pith-review-sha256/v1","hashed_at":"2026-11-20T09:43:44.451306+00:00","fields":["id","arxiv_id","paper_version","created_at","pipeline_version","model_set","run_record","input_kind","input_chars","input_total_chars","passes_missing","verdict","confidence","novelty_score","correctness_risk","formal_verification","parameter_count","discovery_kind","one_line_summary","lead","skeptic_critique","referee_report","author_rebuttal","desk_editor","pith_extraction","objection_audit","published_standing","published_confidence"],"matches_current":true},"falsifier":"If an independent evaluation using a different judge (e.g., GPT-4o as in the original FDB v3 benchmark) or human annotation showed that Duopra's Pass@1 gain over RV-Audio + Harness is substantially smaller (e.g., less than 20 percentage points), the central claim that initiation is the primary bottleneck and that Duopra recovers unserved requests would be weakened.","supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":"tex"},"url":"https://pith.science/accuracy"},"review_version":1}