Artificial Intelligence · arXiv:2610.11299 v1 · submitted 8 Oct 2026 · 19 references
DuplexAgent-RSI: Recursive Harness Improvement for Full-Duplex Voice Agent Collaboration
The paper says
“Voice agents are converging on a collaboration pattern: a full-duplex interaction model stays on the live channel as the entry to the conversation, while search, reasoning, and coding are handled through asynchronous delegation.”
more of the abstract
A duplex model supports continuous listening and speaking, but complex reasoning and tool use may exceed its capabilities. A coding agent can plan and execute extended tasks, but its sequential interface is a poor fit for live conversation. Combining them requires a harness that coordinates task acceptance, progress, cancellation, replacement, and result delivery while keeping the conversation responsive. Existing harnesses often rely on coupled heuristics, making them difficult to improve systematically from evidence. We present DuplexAgent, a full-duplex collaboration system whose harness expresses this workflow as six editable modules, and Duplex-Harness-RSI, a closed loop that revises them from interaction traces. A simulator automatically generates timed test conversations, runs the system, and produces failure traces that identify the collaboration modules requiring repair. Reasoning LLMs and coding agents in the delegation pool also serve the improvement loop: the Exam Planner selects the next tests from observed weaknesses and the repair archive, and the Harness Editor proposes targeted module changes. The capabilities that serve the user thus also improve the system's coordination. Experiments on intelligence, agentic, and duplex benchmarks show that DuplexAgent combines continuous interaction with difficult reasoning and complex task execution, achieving stronger spoken-knowledge and executable-tool scores than the compared delegated systems while maintaining strong interruption response. A harness ablation further shows that this modular, verifiable loop outperforms the initial harness and repeated editing that lacks its diagnosis and repair archive.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 11 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up. A well-engineered system for improving voice agent coordination via a recursive loop; the failure attribution heuristic is the obvious weak point, but the overall results are believable. This line is wrong
2 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 19 references say.
The paper says, at §3.3, Table 2
The review did not quote this passage.
Pith says: Objection 1 of 2 · major
The failure attribution heuristic is not independently validated. The Diagnoser maps each failure to a module using a fixed priority and a default mapping (Table 2). The paper argues that the overall score improvement justifies the mapping, but this is indirect: improvement could also arise from the editor's exploration, the planner's test selection, or the protection gate.
more of objection 1
A small human evaluation on representative cases or a sensitivity analysis where attribution is perturbed would strengthen the claim that the loop benefits from targeted diagnosis.
The paper says, at §5.3, Table 6
The review did not quote this passage.
Pith says: Objection 2 of 2 · major
The local evidence for each repair (e.g., 'Extra calls reduced by 50%') is reported on the exam used to test that specific repair, but the fixed-core score combines all retained repairs. The paper should discuss whether the individual module improvements are additive or whether there are interaction effects, or at least acknowledge that the local gains may not translate linearly to the composite score.
5 minor objections
The paper says, at Abstract
Pith says: minor objection 1 of 5
The phrase 'continuous interaction with difficult reasoning' is slightly awkward; consider rephrasing to 'continuous interaction alongside demanding reasoning' or similar.
The paper says, at §3.3, Eq. (3)
Pith says: minor objection 2 of 5
The notation φ(τ) for the failure signature is introduced but not used later in the text. Consider removing it if it is not needed for the exposition.
show 3 more
The paper says, at Figure 1
Pith says: minor objection 3 of 5
The blue/gray color scheme for the bars may be difficult to distinguish in grayscale print. Consider adding hatching patterns or labels directly on the bars.
The paper says, at Tables 4 and 5
Pith says: minor objection 4 of 5
The marks ∗ and † are defined in the table captions but may be missed by readers. Consider adding an explicit note in each table that ∗ denotes VoiceChat attachment and † denotes Qwen-Realtime attachment.
The paper says, at §4, Table 3
Pith says: minor objection 5 of 5
The configuration details are helpful. Please add a brief estimate of the computational cost (e.g., total GPU-hours for a campaign) to aid reproducibility and resource planning.
What would settle it
Sample 100 failure traces from the VoiceChat RSI campaign (stratified by round and assigned module). Have two independent human annotators, given the trace and the six-module taxonomy, assign the module responsible for each failure. Compute Cohen's kappa between annotators and between each annotator and the automatic attribution (Table 2). Report accuracy of the automatic method against the consensus label. If accuracy is below 75% or kappa below 0.6, the attribution heuristic is unreliable, and the loop’s success may stem from factors other than targeted diagnosis.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 11 Oct 2026
- Reader holds up
- Skeptic agrees
- Editor wrote the verdict above
- Referee minor revision: 2 major objections, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
The paper delivers a practical system for full-duplex voice agent collaboration with a recursive harness improvement loop. The modular decomposition into six editable harness modules (Policy, Intake, Hold, Worker, Deliver, Memory) is sensible and the closed-loop improvement that reuses the delegation pool for diagnosis and editing is novel in this space. The evaluation is thorough: external benchmarks (BFCL, FDB-v3, OBQA, MMSU) show substantial gains over baseline systems, and the ablation comparing one-shot, sequential, and recursive editing on the fixed core is convincing. The improvements on both VoiceChat and Qwen-Realtime attachments are consistent and the version-lineage plot (Figure 7) shows genuine accumulation.
The biggest soft spot is exactly what the reader flags: the failure attribution heuristic in Table 2 is not independently validated. The Diagnoser uses a fixed priority order and default mapping without ground-truth labels. The paper argues that overall score improvement justifies the mapping, but that is a post-hoc argument—it's possible the editor's exploration or the protection gate's filtering is doing the work. That said, the controlled ablation against one-shot and sequential editing suggests the targeted diagnosis adds value, so this is a minor concern rather than a load-bearing flaw. I'd like to see a future validation study, but it doesn't undermine the current claims.
Other soft spots: the results depend on proprietary models (DeepSeek-v4-flash, Codex, Qwen-Realtime), so independent reproduction is limited. The protection gate may not fully prevent overfitting if the test distribution drifts, though the fixed core helps. The paper acknowledges that the loop stops at the harness and does not improve the interaction model's perception or timing—an honest limitation.
Overall, this is a solid systems paper. The architecture is clean, the empirical evidence is strong given the constraints, and the central claim holds up. The failure attribution heuristic deserves scrutiny but is not a dealbreaker.
Who is this for? Researchers working on voice agents, automated agent design, and recursive self-improvement. It deserves a serious referee and should be published after minor revisions that add a discussion of attribution validation or a plan for it.
I would cite this if I were working on voice agent coordination.
Recommendation: send to peer review. The paper is well-executed and the claims are supported.
The strongest objection
Duplex-Harness-RSI's iterative improvement depends on the Diagnoser assigning each failure to one of six modules using a fixed priority and default mapping (Table 2, §3.3). This heuristic is not validated against ground-truth labels. If attribution is wrong, the Exam Planner may select tests that miss the actual weak module, and the Harness Editor may edit an irrelevant module, producing ineffective or harmful changes. The paper argues that overall score improvement justifies the mapping, but this is circular: improvement could arise from the editor's exploration, test selection, or gate filtering rather than accurate attribution. Without independent validation, the central claim that the loop benefits from targeted diagnosis is unsubstantiated. The reader's weakest assumption also identifies this, confirming its centrality.
The reader's weakest assumption was the accuracy of the failure attribution, and our analysis confirms this as the most load-bearing concern. The entire RSI loop—Exam Planner, Harness Editor, and gate decisions—depends on mapping failures to the correct module. Without independent validation, the paper's evidence that the loop improves scores is consistent with correct attribution but does not rule out alternative explanations (e.g., broad or exploratory edits passing the gates). We propose a concrete human-annotation check to directly measure attribution accuracy. The reader already acknowledged this vulnerability and still accepted the paper with moderate confidence, and we do not find a reason to change that verdict because the overall improvement results are strong and the paper provides some indirect validation via sequential editing baselines. However, the lack of direct validation remains a notable gap that future work or an additional experiment could address. Hence verdict unchanged.
The referee's report
If the results hold, this work makes a strong contribution by showing that a modular harness for full-duplex delegation can be systematically improved from interaction traces using the same models that serve user requests. The separation of the harness into six editable modules and the recursive improvement loop (Exam Planner + Harness Editor + simulator) are novel and practically valuable for building spoken-agent systems that combine real-time conversation with long-horizon reasoning and tool use. The experimental design uses external benchmarks and a controlled ablation, and the code is released, supporting reproducibility.
Pith's reading of the paper (a machine's summary, not the authors')
This paper argues that the coordination 'harness' between a full-duplex voice interaction model and a delegation pool of reasoning/coding agents can be systematically improved by decomposing it into six editable modules (Policy, Intake, Hold, Worker, Deliver, Memory) and running a recursive self-improvement loop (Duplex-Harness-RSI). The loop uses the same delegation pool to act as an Exam Planner and Harness Editor: it generates timed test conversations, runs them, attributes failures to specific modules, and proposes targeted repairs. The authors claim that this closed-loop approach significantly outperforms one-shot or sequential editing baselines, raising collaboration scores on fixed evaluation sets from roughly 55% to 85% on one interaction model and from 47% to 85% on another. If correct, the paper shows how to turn coordination failures into measurable, repairable components, making full-duplex voice agents simultaneously more responsive and more capable at complex reasoning and tool use.
What carries the argument. The key machinery is the decomposition of the collaboration harness into six editable modules (Policy, Intake, Hold, Worker, Deliver, Memory), each with a defined decision and editable surface, and the recursive self-improvement loop (Duplex-Harness-RSI) that reuses the delegation pool. A simulator generates timed test conversations by composing request types (19 types) with interaction patterns (22 patterns). The Observer captures a trace, the Diagnoser attributes failures to a module using a fixed priority order and default mapping, the Exam Planner (a reasoning LLM) adjusts test sampling based on these attributions and the repair archive, and the Harness Editor (a coding agent) proposes module changes. Candidates are evaluated on paired runs with a development gate (more wins than losses) and a protection gate (non-degradation on a held-out set). The fixed core of test cases is repeated across rounds to measure progress.
Core claim. The central discovery is that the coordination harness for full-duplex voice agent collaboration can be expressed as six editable modules (Policy, Intake, Hold, Worker, Deliver, Memory), and that a recursive improvement loop (Duplex-Harness-RSI) can diagnose failures at the module level using a simulator and then repair those modules using the same reasoning and coding agents that serve user requests. The loop includes an Exam Planner that adapts test coverage based on observed weaknesses and a Harness Editor that proposes targeted module changes. The authors demonstrate that this approach raises collaboration scores on a fixed core of test conversations from 54.6 to 85.3 (VoiceChat) and from 47.1 to 84.8 (Qwen-Realtime), substantially outperforming baselines that pool or sequentially replay the same test material without the diagnosis and repair archive. The retained repairs improve spoken-knowledge accuracy (e.g., MMSU from 36.1% to 80.5% on VoiceChat) and executable tool use (e.g., BFCL average from 53.5% to 75.8%) while maintaining strong interruption response timing.
The premise it rests on. The system's ability to correctly attribute a coordination failure to a specific editable module depends on a fixed priority order and a default mapping (Table 2) that is not independently validated; if the attribution is wrong, the Harness Editor may modify the wrong module and the Exam Planner may select irrelevant tests, yet the overall improvement suggests the mapping is usually sufficient for the loop to make progress.
If the paper is right (Pith's inferences, not the authors')
If correct, this method provides a template for systematically improving any multi-agent coordination system where the harness logic can be modularized and failures diagnosed: the same loop could be applied to customer service bots, collaborative coding agents, or robotic task delegation.
The approach suggests that the reasoning and coding agents used for user tasks can double as improvement machinery without separate training, potentially reducing the cost of maintaining complex voice interfaces over time.
The modular harness enables transparent debugging: a user or developer can see which module caused a coordination failure and track which repairs were accepted or rejected across rounds, building a knowledge base of effective coordination patterns.
The paper implies that the limits of harness improvement are set by the interaction model's perception and timing (e.g., overlapped speech detection), which remains outside the editable modules; future work could extend the loop to train the interaction model itself using the same simulator, as the authors suggest in the discussion.
An unstated implication is that this modular, trace-driven improvement loop could replace hand-tuned heuristics in many current voice-agent deployments, making coordination robustness a function of available compute and test coverage rather than engineering intuition.
The result suggests a division of labor: the interaction model handles smooth conversation (backchannels, barge-in, timing) while the harness handles task identity, state, and delivery. This separation implies that improvements to the interaction model and harness are orthogonal and can be pursued in parallel.
A testable extension not pursued in the paper: the same diagnosis and repair loop could be applied to systems with multiple interacting harnesses (e.g., a team of voice agents), where failures in inter-agent coordination are attributed to modules of a shared harness.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Voice agents are converging on a collaboration pattern: a full-duplex interaction model stays on the live channel as the entry to the conversation, while search, reasoning, and coding are handled through asynchronous delegation. A duplex model supports continuous listening and speaking, but complex reasoning and tool use may exceed its capabilities. A coding agent can plan and execute extended tasks, but its sequential interface is a poor fit for live conversation. Combining them requires a harness that coordinates task acceptance, progress, cancellation, replacement, and result delivery while keeping the conversation responsive. Existing harnesses often rely on coupled heuristics, making them difficult to improve systematically from evidence. We present DuplexAgent, a full-duplex collaboration system whose harness expresses this workflow as six editable modules, and Duplex-Harness-RSI, a closed loop that revises them from interaction traces. A simulator automatically generates timed test conversations, runs the system, and produces failure traces that identify the collaboration modules requiring repair. Reasoning LLMs and coding agents in the delegation pool also serve the improvement loop: the Exam Planner selects the next tests from observed weaknesses and the repair archive, and the Harness Editor proposes targeted module changes. The capabilities that serve the user thus also improve the system's coordination. Experiments on intelligence, agentic, and duplex benchmarks show that DuplexAgent combines continuous interaction with difficult reasoning and complex task execution, achieving stronger spoken-knowledge and executable-tool scores than the compared delegated systems while maintaining strong interruption response. A harness ablation further shows that this modular, verifiable loop outperforms the initial harness and repeated editing that lacks its diagnosis and repair archive.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-09
The authors' main claim, as the classifier picked it out 1.00
“Experiments on intelligence, agentic, and duplex benchmarks show that DuplexAgent combines continuous interaction with difficult reasoning and complex task execution, achieving stronger spoken-knowledge and executable-tool scores than the compared delegated systems while maintaining strong interruption response.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.15 · human participants 0.08 · controlled comparison 0.89 · randomized assignment 0.02 · states quantitative result 0.07 · compares with prior results 0.96 · mentions code or data availability 0.04
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Voice agents are converging on a collaboration pattern: a full-duplex interaction model stays on the live channel as the entry to the conversation, while search, reasoning, and coding are handled through asynchronous delegation. | background 1.00 | 0.06 | 0.04 | asserted 0.99 |
| A duplex model supports continuous listening and speaking, but complex reasoning and tool use may exceed its capabilities. | background 0.91 | 0.08 | 0.02 | hedged possibility 0.91 |
| A coding agent can plan and execute extended tasks, but its sequential interface is a poor fit for live conversation. | background 0.93 | 0.08 | 0.02 | asserted 1.00 |
| Combining them requires a harness that coordinates task acceptance, progress, cancellation, replacement, and result delivery while keeping the conversation responsive. | background 0.70 | 0.07 | 0.04 | asserted 0.99 |
| Existing harnesses often rely on coupled heuristics, making them difficult to improve systematically from evidence. | gap 0.77 | 0.07 | 0.02 | asserted 0.94 |
| We present DuplexAgent, a full-duplex collaboration system whose harness expresses this workflow as six editable modules, and Duplex-Harness-RSI, a closed loop that revises them from interaction traces. | contribution 0.99 | 0.94 | 0.15 | asserted 1.00 |
| A simulator automatically generates timed test conversations, runs the system, and produces failure traces that identify the collaboration modules requiring repair. | method 0.99 | 0.83 | 0.03 | asserted 1.00 |
| Reasoning LLMs and coding agents in the delegation pool also serve the improvement loop: the Exam Planner selects the next tests from observed weaknesses and the repair archive, and the Harness Editor proposes targeted module changes. | method 0.96 | 0.82 | 0.05 | asserted 1.00 |
| The capabilities that serve the user thus also improve the system's coordination. | interpretation 0.96 | 0.42 | 0.09 | asserted 1.00 |
| Experiments on intelligence, agentic, and duplex benchmarks show that DuplexAgent combines continuous interaction with difficult reasoning and complex task execution, achieving stronger spoken-knowledge and executable-tool scores than the compared delegated systems while maintaining strong interruption response. | result 1.00 | 0.96 | 0.79 | asserted 1.00 |
| A harness ablation further shows that this modular, verifiable loop outperforms the initial harness and repeated editing that lacks its diagnosis and repair archive. | result 1.00 | 0.96 | 0.16 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.97
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-09: 1 request, 8,259 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (7)
Figures are the authors', reproduced from the paper.
References (19)
- [1] Ant Group (Venus Team) and Tsinghua University. Realtime-venus: A full-duplex interaction system with asynchronous delegation.arXiv preprint arXiv:2609.13814, on PithPith's reviewarXiv
- [3] URL https://arxiv.org/abs/2609.21967. Model card:https://huggingface.co/nvidia/NVIDIA-NemotronLabs-VoiceChat-11B. Junjie Chen, Yao Hu, Junjie Li, Kangyue Li, Kun Liu, Wenpeng Li, Xu Li, Ziyuan Li, Feiyu Shen, Xu Tang, et al. FireRedChat: A pluggable, full-duplex voice interact... on PithPith's reviewarXiv
- [6] Junbo Cui et al. MiniCPM-o 4.5: Towards real-time full-duplex omni-modal interaction.arXiv preprint arXiv:2604.27393, on PithPith's reviewarXiv
- [7] Alexandre Défossez, Laurent Mazaré, Manu Orsini, Amélie Royer, Patrick Pérez, Hervé Jégou, Edouard Grave, and Neil Zeghidour. Moshi: A speech-text foundation model for real-time dialogue.arXiv preprint arXiv:2410.00037, on PithPith's reviewarXiv
- [9] URL https://arxiv.org/ abs/2609.25195. Zhang He, Wenqian Cui, Haoning Xu, Xiao-Hui Li, Lei Zhu, Haoli Bai, Ma Shaohua, and Irwin King. MTR-DuplexBench: Towards a comprehensive evaluation of multi-round conversations for full-duplex speech language models. InFindings of ACL, pp... on PithPith's reviewarXiv
- [10] Yoonho Lee, Roshen Nair, Qizheng Zhang, Kangwook Lee, Omar Khattab, and Chelsea Finn. Meta- harness: End-to-end optimization of model harnesses.arXiv preprint arXiv:2603.28052, on PithPith's reviewarXiv
- [11] Guan-Ting Lin, Shih-Yun Shan Kuan, Qirui Wang, Jiachen Lian, Tingle Li, Shinji Watanabe, and Hung-yi Lee. Full-duplex-bench v1.5: Evaluating overlap handling for full-duplex speech models.arXiv preprint arXiv:2507.23159, 2025a. Guan-Ting Lin, Jiachen Lian, Tingle Li, Qirui Wan... on PithPith's reviewarXiv
- [12] Yueqian Lin, Zhengmian Hu, Jayakumar Subramanian, Qinsi Wang, Nikos Vlassis, Yiran Chen, et al. AsyncVoice agent: Real-time explanation for LLM planning and reasoning. InProc. ASRU, pp. 1–4, 2025c. Zhanxun Liu, Yifan Duan, Mengmeng Wang, Pengchao Feng, Haotian Zhang, Xiaoyu Xi... on PithPith's reviewarXiv
- [13] Xufang Luo, Yuge Zhang, Zhiyuan He, Zilong Wang, Siyun Zhao, Dongsheng Li, Luna K Qiu, and Yuqing Yang. Agent lightning: Train any AI agents with reinforcement learning.arXiv preprint arXiv:2508.03680, on PithPith's reviewarXiv
- [14] 15 T echnical Report 16 Orantqing, Shengpeng Ji, Junlong Tong, Jialong Zuo, et al. Multimodal duplex interaction agent.arXiv preprint arXiv:2609.08977, on PithPith's reviewarXiv
- [15] URLhttps://arxiv.org/abs/2603.13686. Rajarshi Roy, Jonathan Raiman, Sang-gil Lee, Teodor-Dumitru Ene, Robert Kirby, Sungwon Kim, Jaehyeon Kim, and Bryan Catanzaro. PersonaPlex: Voice and role control for full duplex conversational speech models. InProc. ICASSP, pp. 16137–16141, on PitharXiv
- [16] Zihan Tan, Leixin Sun, Zitong Shi, Yitao Liu, Jiajun Wu, et al. MetaRSI / RSI 2: A meta-recursive self- improving system for recursive self-improving systems themselves.arXiv preprint arXiv:2609.06396, on PithPith's reviewarXiv
- [17] Chengyou Wang, Hongfei Xue, Guojian Li, Zhixian Zhao, Shuiyuan Wang, Shuai Wang, Xin Xu, Hui Bu, and Lei Xie. Full-duplex interaction in spoken dialogue systems: A comprehensive study from the ICASSP 2026 HumDial challenge.arXiv preprint arXiv:2604.21406, on PithPith's reviewarXiv
- [18] Xiong Wang, Yangze Li, Chaoyou Fu, Yunhang Shen, Lei Xie, Ke Li, Xing Sun, and Long Ma. Freeze- omni: A smart and low latency speech-to-speech dialogue model with frozen LLM.arXiv preprint arXiv:2411.00774, on PithPith's reviewarXiv
- [19] Siwei Wu, Jincheng Ren, Yizhi Li, Haau-Sing Li, Chengran Yang, Yuxuan Zhang, Weicheng Gu, Jian Yang, Riza Batista-Navarro, Chuanyi Zhang, Xianglong Liu, Ming Zhou, Bryan Dai, and Chenghua Lin. ModularRSI: Modular and generalizable recursive harness self-improvement.arXiv prepr... on PithPith's reviewarXiv
- [21] Hangfan Zhang, Shao Zhang, Kangcong Li, Chen Zhang, Yang Chen, Yiqun Zhang, Lei Bai, and Shuyue Hu. Self-harness: Harnesses that improve themselves.arXiv preprint arXiv:2606.09498, on PithPith's reviewarXiv
- [2023] Wenyi Yu, Siyin Wang, Xiaoyu Yang, Xianzhao Chen, Xiaohai Tian, Jun Zhang, Guangzhi Sun, Lu Lu, Yuxuan Wang, and Chao Zhang. SALMONN-omni: A codec-free LLM for full-duplex speech under- standing and generation.arXiv preprint arXiv:2411.18138, on PithPith's reviewarXiv
- [2024] Chung-Ming Chien, Manu Orsini, Eugene Kharitonov, Neil Zeghidour, Karen Livescu, and Alexandre Défossez. MoshiRAG: Asynchronous knowledge retrieval for full-duplex speech language models. arXiv preprint arXiv:2604.12928, on PithPith's reviewarXiv
- [2025] Yiming Chen, Xianghu Yue, Chen Zhang, Xiaoxue Gao, Robby T. Tan, and Haizhou Li. VoiceBench: Benchmarking LLM-based voice assistants.arXiv preprint arXiv:2410.17196, on PithPith's reviewarXiv
Cited by 3 Pith papers
Reviewed papers on Pith that cite this one. The full citation record
- Decode Before You Hear: Speculative Listening for Chunk-Interleaved Full-Duplex Speech Language Models cs.SD · Nov 2026
- GraphSeal: Graph-Augmented Speech Deepfake Detection via Temporal-Spectral Artifact Propagation cs.SD · Nov 2026
- Duopra: From Full-Duplex Conversation to Full-Duplex Agency cs.AI · Nov 2026
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment