Paper deep dive
DuplexWorld: Can voice agents help you get through the day?
Aryan Vijay Bhosale, Harshit Rajgarhia, Akhil Pothanapalli, Asif Shaik, Abhishek Mukherji, Dinesh Manocha
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 8/16/2026, 3:12:33 AM
Summary
The paper introduces DuplexWorld, a benchmark for evaluating speech-to-speech (S2S) voice agents across six diverse worlds (banking, insurance, travel, healthcare, logistics, and Pathfinding). It addresses gaps in existing benchmarks by testing holistic capabilities including agentic tool calling, conversational dynamics, and speech naturalness. The evaluation covers 156 scenarios and 3,825 conversations, revealing that even top-performing agents have significant room for improvement in task completion, turn-taking, and audio quality.
Entities (14)
Relation Signals (13)
DuplexWorld → evaluatesmetric → pass@1
confidence 95% · Through extensive evaluation ... we show that even the best voice agents leave substantial room for improvement on all 3 axes (Pass@1: 0.490...)
DuplexWorld → evaluatesmetric → Turn-taking
confidence 95% · Through extensive evaluation ... we show that even the best voice agents leave substantial room for improvement on all 3 axes (... turn-taking: 0.653...)
DuplexWorld → evaluatesmetric → DNSMOS
confidence 95% · Through extensive evaluation ... we show that even the best voice agents leave substantial room for improvement on all 3 axes (... DNSMOS: 3.378).
DuplexWorld → includesdomain → Healthcare
confidence 95% · DuplexWorld introduces six worlds where voice agents are especially useful: ... healthcare ...
DuplexWorld → includesdomain → Logistics
confidence 95% · DuplexWorld introduces six worlds where voice agents are especially useful: ... logistics ...
DuplexWorld → includesdomain → Pathfinding
confidence 95% · DuplexWorld introduces six worlds where voice agents are especially useful: ... and Pathfinding.
DuplexWorld → includesdomain → Banking
confidence 95% · DuplexWorld introduces six worlds where voice agents are especially useful: banking...
DuplexWorld → →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Speech-to-speech (S2S) voice agents are increasingly being incorporated into enterprise for customer care and as daily companions for consumers owing to the ease of the conversational modality over text. However, existing benchmarks fail to holistically evaluate voice agents along axes that really matter and are shaped as tests of agentic tool calling against a database. We believe they fail to adequately account for the diversity of conversational dialogue that mundane activities introduce and further, never test how faithfully an agent can assist on tasks that move beyond database manipulation. To tackle this DuplexWorld introduces six worlds where voice agents are especially useful: banking, insurance, travel, healthcare and logistics, and Pathfinding. Agents are evaluated on eleven different types of conversations across 156 scenarios (350+ hours of conversation), each testing conversational and analytical capability to varying degrees. Through extensive evaluation comprising agentic, conversational and speech-naturalness metrics, we show that even the best voice agents leave substantial room for improvement on all 3 axes (Pass@1: 0.490, turn-taking: 0.653, DNSMOS: 3.378). We perform extensive analysis on agentic v conversational performance, world- and conversation type-wise performance, failure modes exploring the explore v exploit lens for Pathfinding conversations and voice agent reliability over all six worlds.
Tags
Links
- Source: https://arxiv.org/abs/2608.10716v1
- Canonical: https://arxiv.org/abs/2608.10716v1
Trouble viewing inline? Open PDF directly →
Full Text
96,016 characters extracted from source content.
Expand or collapse full text
DuplexWorld: Can voice agents help you get through the day? Aryan Vijay Bhosale1,2∗, Harshit Rajgarhia1∗, Akhil Pothanapalli1∗, Asif Shaik1, Abhishek Mukherji1†, Dinesh Manocha2† 1Centific Global Solutions Inc. 2University of Maryland aryan.bhosale, harshit.rajgarhia@centific.com https://duplexworld.github.io Abstract Speech-to-speech (S2S) voice agents are increasingly being incorporated into enterprise for customer care and as daily companions for consumers owing to the ease of the conversational modality over text. However, existing benchmarks fail to holistically evaluate voice agents along axes that really matter and are shaped as tests of agentic tool calling against a database. We believe they fail to adequately account for the diversity of conversational dialogue that mundane activities introduce and further, never test how faithfully an agent can assist on tasks that move beyond database manipulation. To tackle this DuplexWorld introduces six worlds where voice agents are especially useful: banking, insurance, travel, healthcare and logistics, and Pathfinding. Agents are evaluated on eleven different types of conversations across 156 scenarios (350++ hours of conversation), each testing conversational and analytical capability to varying degrees. Through extensive evaluation comprising agentic, conversational and speech-naturalness metrics, we show that even the best voice agents leave substantial room for improvement on all 3 axes (Pass@1Pass@1: 0.490, turn-taking: 0.653, DNSMOS: 3.378). We perform extensive analysis on agentic v conversational performance, world- and conversation type-wise performance, failure modes exploring the explore v exploit lens for Pathfinding conversations and voice agent reliability over all six worlds. DuplexWorld: Can voice agents help you get through the day? Aryan Vijay Bhosale1,2∗, Harshit Rajgarhia1∗, Akhil Pothanapalli1∗, Asif Shaik1, Abhishek Mukherji1†, Dinesh Manocha2† 1Centific Global Solutions Inc. 2University of Maryland aryan.bhosale, harshit.rajgarhia@centific.com https://duplexworld.github.io †footnotetext: Work done during an internship at Centific. ∗Equal contribution. †Equal advising. 1 Introduction Figure 1: Reliability curves for voice-agent performance over all six worlds. Dashed: Pass@kPass@k, at least one pass in k attempts; solid: PasskPass^k, all k. Voice Think Fast leads both. (a) (b) Figure 2: DuplexWorld at a glance. (a) The six worlds of an ordinary day, the authored scenarios (one wedge per world), and conversation-time densities with hours of speech per world. (b) Six of the eleven conversation types (Section 3.2), each with a verbatim passing excerpt from the scored corpus; far right, the two Pathfinding-only types with the walked route against the ideal routes. Failing analogues of all eight tiles are in Figure 6 (appendix). Speech-to-speech (S2S) voice agents capable of simultaneously understanding and generating audio increasingly exhibit remarkable agentic capabilities. Voice offers an expressive and convenient interface, making such agents preferable to text-based agents. In enterprise they answer customer-care lines and help with re-booking flights, while for consumers they act as a constant companion, helping them find their way around a new city or run daily errands. Each of these is a conversation with nuanced conversational dynamics and analytical demands: reconciling records, guiding users through busy streets, and doing both while contending with muffled speech, background noise and an emotionally diverse range of users. Full-duplex voice agent research has proliferated across academia and industry. Benchmarks like τ-Voice showed that voice agents retain only 30–45% of text-agent capability on identical grounded tasks under realistic audio (Ray et al., 2026), and EVA-Bench showed that no system is simultaneously good at task accuracy and conversational experience (Bogavelli et al., 2026). Full-Duplex-Bench progressed from static turn-taking probes to real disfluent speech with chained tool calls (Lin et al., 2025b, 2026b, a, 2026a), while parallel lines isolate interruption and repair (Modi et al., 2026; Salimi et al., 2026; Lu et al., 2026). While the difficulty and framing of these tasks and evaluations were justified at the time, the voice agents of today are far more capable and deserve benchmarks that keep up with their rapid development. While prior benchmarks pursued coverage across domains in their own ways, they failed to question the base premise on which voice agents are built: their ability to integrate seamlessly into daily life. To address these issues, we introduce DuplexWorld, a benchmark for unified, holistic evaluation of voice agents across conversational and analytical tasks emerging from six diverse worlds and eleven conversation types. The banking, insurance, travel, healthcare and logistics worlds consolidate and enhance evaluation along domains similar to those explored in existing benchmarks, while Pathfinding introduces a new frontier for testing voice-agentic capability. Our contributions are: • Six worlds spanning 156 scenarios and eleven conversation types, expanding the gamut of voice-agentic evaluation. The first navigation world for full-duplex voice agents joins five enterprise worlds, with nine analytical interaction shapes shared across the enterprise worlds and two that only navigation elicits: the broadest coverage of any voice-agent benchmark (156 authored scenarios, 3,825 scored conversations; Sections 3.1–3.2). • One unified evaluation suite across all of it. Twelve metrics in three pillars run under one harness and one configuration in every world: to our knowledge the first single suite spanning this breadth of domains and demands (Section 4.3). • Evidence that the three capabilities do not travel together. Pass@1Pass@1 for one system spans 0.200–0.674 across worlds differing only in subject matter while the ranking barely moves; the best conversationalists are not the best task-completers; acoustic quality does not predict competence; the under-effort share π−π^- predicts success at ρ=−0.85ρ=-0.85; and read through an explore–exploit lens, the heaviest explorers in Pathfinding arrive least (Section 5). 2 Related Work full duplex dynamics agentic task naturalness reliability PasskPass^k sim. valid. harness sens. nav. world τ-bench (2024) −- −- ✓ −- ✓ −- −- −- τ2τ^2-bench (2025) −- −- ✓ −- ✓ −- −- −- FDB v1–v2 (2025b) ✓ ✓ −- −- −- −- −- −- Talking Turns (2025) ✓ ✓ −- −- −- −- −- −- HumDial (2026) ✓ ✓ −- ✓ −- −- −- −- EchoChain (2026) ✓ ✓ ✓ −- −- −- −- −- IHBench (2026) ✓ ✓ ✓ −- −- −- −- −- FDB v3 (2026a) ✓ ✓ ✓ −- −- −- −- −- τ-Voice (2026) ✓ −- ✓ −- ✓ −- −- −- EVA-Bench (2026) ✓ ✓ ✓ −- ✓ ✓ −- −- DuplexWorld (ours) ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Table 1: Comparison of DuplexWorld with prior voice-agent and text-agent benchmarks. The three grouped columns are the metric pillars of Section 4.3. Reliability: a multi-run repeatability metric (PasskPass^k); sim. valid.: automated user-simulator validation; harness sens.: quantified sensitivity to harness configuration; nav. world: a navigation world. Task-oriented agent benchmarks with verifiable outcomes. τ-bench introduced database-state grading and PasskPass^k, the probability that all k runs of a scenario succeed, for text agents in retail and airline customer service (Yao et al., 2024), and τ2τ^2-bench added a telecom domain in which the user also holds tools, reporting an 18–25 point Pass@1Pass@1 drop when agents move from acting alone to guiding a user (Barres et al., 2025). These benchmarks fixed the right question, does the world end up in the right state, but asked it of text agents. DuplexWorld keeps their discipline (state grading, gold actions, PasskPass^k) and carries it into full-duplex voice, across six worlds instead of two or three domains. Full-duplex conversational dynamics. A second line asks whether a system can hold the floor: Full-Duplex-Bench defined automatic turn-taking metrics and its successors added overlap handling and a live automated examiner (Lin et al., 2025b, 2026b, a); Talking Turns judges turn-taking against human conversation (Arora et al., 2025); the ICASSP 2026 HumDial challenge released dual-channel real human dialogues (Zhao et al., 2026; Wang et al., 2026); and a recent survey organises the architectural space by where the duplex decision sits in the model stack (Lu et al., 2026). These were the right first tests, and the current generation of commercial realtime systems has largely grown past them; our results sharpen the point by showing that dynamics metrics alone are winnable by silence (Section 5). FDB-v3, Talking Turns and HumDial evaluate on real human audio, and we do not (Limitations). DuplexWorld therefore keeps conversational dynamics as one pillar of a unified suite, paired with capability metrics that silence cannot win. Grounded voice-agent benchmarks. The newest line puts stronger voice agents on verifiable tasks: τ-Voice with a tick-based orchestrator over 278 grounded tasks (Ray et al., 2026), EVA-Bench with validation-gated simulation over 213 enterprise scenarios (Bogavelli et al., 2026), FDB-v3 with real disfluent speech and chained tool calls (Lin et al., 2026a), and the interruption-recovery suites EchoChain and IHBench (Modi et al., 2026; Salimi et al., 2026). Harder tests for better agents, yet in every one of them the agent actuates, a record system adjudicates, and nothing happens between utterances, and the interruption work varies only the user’s intent (Table 1). DuplexWorld spans that setting and a navigation world under one harness and one metric suite, and it is hard even for this new generation: no system in our corpus exceeds Pass@1Pass@1 of 0.674 in any single world, or 0.533 in the world that moves on its own clock. 3 The DuplexWorld Benchmark DuplexWorld pairs a voice agent with a simulated user in six worlds of an ordinary day for 156 authored scenarios, of which 144 carry the numbers in this paper. Every world is conversational and every world is analytical: each couples live full-duplex speech to a verifiable task whose correctness the transcript alone cannot decide. Every world is scored on one twelve-metric suite spanning three pillars (conversational dynamics, agentic capability, naturalness), and no composite is ever formed (one word per level of the hierarchy; Appendix A). Every scenario was authored by us over multiple rounds of experimentation, including trials with text-mode language models; the task-design discipline follows τ-bench and EVA-Bench (Yao et al., 2024; Bogavelli et al., 2026). 3.1 Task formulation An episode pairs one voice agent with one simulated user in one world for a bounded number of ticks. The world exposes a state s, a tool interface and a graded end condition. The agent perceives only audio and tool returns, and it emits only audio and tool calls. When the episode ends, the terminal state is scored against a gold state s⋆s authored with the scenario, so success is a property of the world rather than of the transcript. State advances through whichever effectors the world grants, st+1=T(st,at,ut)s_t+1=T(s_t,a_t,u_t). In Banking, Insurance, Travel, Healthcare and Logistics the records move only through the agent’s tool calls ata_t. In Pathfinding the world moves only through the walker’s physical actions utu_t, which the agent can influence through language alone, so state there can change while both parties are silent, and an utterance planned at tick t can be wrong by the time it lands at tick t+kt+k. The five enterprise worlds share one premise. A caller reaches an institution’s telephone line after self-service has already refused them. The agent owns the call and may not route away. The governing clause corpus is withheld from the system prompt, so knowing when to look something up is itself under test. The agent cannot see the caller’s screen or surroundings, and credentials travel by voice over a telephone codec, a deliberate stressor. What distinguishes the five worlds is what decides correctness (Table 7), and several regimes are enforced in the tool surface itself, which makes the unsafe behaviour unrepresentable rather than discouraged (Appendix D). Pathfinding asks the same question, does the world end up in the right state, but its correctness is perceptual rather than institutional. A voice copilot must walk a pedestrian, one street corner at a time, to a named entrance on a synthetic 8×88× 8 urban grid whose two superblocks defeat naive rectilinear reasoning. The premise is an information asymmetry. The walker holds their real junction and heading but knows nothing of the grid; the copilot holds the complete map and approximate GPS, but no tool returns the walker’s facing, and directions() returns absolute bearings the walker cannot act on. Translating those bearings into “keep going straight” or “turn around”, against the copilot’s own inferred belief about the walker’s facing, is the operative skill, and it mirrors how production routing APIs behave. The walker must be standing at the destination when the call ends, not merely have passed through it; route efficiency enters the reward and the acoustic channel is varied. The full tool surface, belief probe and arrival scoring are in Appendix E, and Appendix F draws a solved and an unsolved run of each Pathfinding type on the map itself. 3.2 Task design Every scenario instantiates one of eleven conversation types (Table 8): a domain-independent interaction shape, filled in by the world’s premise, records, rules and risk tiers. The five enterprise worlds each instantiate types 1–9, three scenarios per type: 27 per world, 135 in total. Pathfinding instantiates single intent (its base routes) together with the two types that only navigation elicits: rerouting, where closures absent from the copilot’s map seal every route it can see and exactly one detour survives, and all-day assistance, where the walker reveals a second destination only on arriving at the first. Each of the three types has three authored scenarios, giving nine campaigned Pathfinding scenarios. Every scenario is run five times per channel over two acoustic channels, so Pathfinding contributes 9×5×2=909× 5× 2=90 conversations per system; the realistic channel’s 45 carry its headline cells (Section 4.2). Together with the 135 enterprise scenarios, these nine complete the 144 scenarios that carry every number in this paper; the remaining twelve of the 156 authored scenarios belong to four Pathfinding variants that are released but not run (Appendix E). Types are shapes, not levels: they are unordered, and parameters that govern difficulty (escalation count, candidate-record count, closure count) live inside a type. Figure 2 shows a verbatim passing excerpt for six of the types; Figure 6 (appendix) shows each type’s characteristic failure from the same corpus. All six worlds share one volatility stressor: the world can change with no notification, discoverably only by re-querying. The three sources of belief–world divergence this exposes, and the different repairs each demands, are catalogued in Appendix D. A benchmark’s honest description also includes what its reward cannot see; ours is audited in Appendix M. 4 Experimental Setup 4.1 Systems Table 2 lists the five commercial realtime speech-to-speech systems evaluated, identified exactly as the run records identify them. All five are served over WebSocket realtime endpoints through one adapter at provider defaults; full serving details are in Appendix C. system API identifier Nova 2 Sonic amazon.nova-2-sonic-v1:0 Gemini-3.1-Flash-Live gemini-3.1-flash-live-preview GPT-Realtime-2.1 gpt-realtime-2.1 GPT-Realtime-2.1-mini gpt-realtime-2.1-mini Grok Voice Think Fast 1.0 xai-realtime Table 2: Systems evaluated, as recorded in the run logs. All are speech-to-speech; elsewhere we write each system as its vendor mark plus the version alone ( 2 Sonic, 3.1-Flash-Live, Realtime-2.1, Realtime-2.1-mini, Voice Think Fast). Access windows are in Appendix C. 4.2 Evaluation strategy Harness. Our harness is built by taking inspiration from τ-Voice and EVA-Bench (Ray et al., 2026; Bogavelli et al., 2026): it follows their tick-based design and simulator instrumentation, and Appendix B states precisely what is inherited and what we added. Simulated time advances in 200 ms ticks; model latency never shifts an event’s position, so slow and fast systems produce comparable traces, and tool results deliver on the next tick so audio never stalls on tool latency. Holding one harness fixed across all six worlds is deliberate: uniformity on the harness side is what lets the eleven conversation types ask novel questions along new axes, and it is what makes the unified evaluation of Section 4.3 possible. Every run pairs the agent under test with a user simulator (gpt-5.6-luna) that writes the caller’s or walker’s words; on the realistic channel a decision model (claude-haiku-4.5) additionally answers “should I interrupt” and “should I backchannel” at a fixed cadence while the agent speaks, and the clean channel disables interruptions and backchannels, so no decision model participates (Table 6). Turn-taking control sits entirely on the user side, and the agent is never told any threshold. Our own additions, and how little of the type taxonomy depends on them, are in Appendix B. Infrastructure failure means rerun and never score; agent non-response is a valid scored end state; simulator goal adherence is measured, and never gates reward (Bogavelli et al., 2026). Task mechanics. Tools in the enterprise worlds are declarative mocks over a seeded record store; each scenario names a gold action list and, where what is said or refused is the point, a natural-language assertion. In Pathfinding, walker motion is maneuver-driven and the copilot’s reads return natural-language strings; the full tool surface is in Appendix E. Conditions and runs. One matched configuration across all six worlds: speech complexity regular (distractors at 0.7/min), step cap 1200 s, wall-clock timeout 1200 s, server VAD threshold 0.2, fixed seed, persona pinned per scenario. Both channels are G.711 μ-law 8 kHz telephony; clean adds no degradation, and realistic adds recorded noise, bursts, frame drops, muffling and speech inserts (Appendix G). The enterprise worlds run realistic-only; Pathfinding runs clean and realistic, and its headline cells in Table 3 use the realistic channel, the same basis as the enterprise worlds, with both channels decomposed in Table 14. Five runs per scenario per system per channel, uniform: 135 conversations per system in each enterprise world, 90 in Pathfinding, 3,825 in total (Appendix P). Uniform n=5n=5 is what makes Pass3Pass^3 well defined. Statistics. Every cell in Table 3 carries a 95% percentile-bootstrap interval (Pass3Pass^3 resampled over scenarios). Overlapping intervals mean a difference is not resolved, and we do not report unresolved differences as rankings. Because the five enterprise worlds share harness, metrics, configuration and type taxonomy, the interval a system’s score spans across them is a measured null, how far a voice-agent number moves when only subject matter changes, and Pathfinding is read against that reference throughout Section 5. 4.3 Metrics Twelve metrics in three pillars, no composite; Table 11 fixes every name, symbol and range, and Appendix H gives the definitions and formulas. Metrics are computed per run and averaged; worlds are pooled by unweighted mean over equal-sized cells; J∙J_ denotes an LLM judge (Table 12). The suite is one evaluation system, not twelve scripts: a single pipeline computes every metric in every world from the same three artefacts (the merged transcript, the simulation tick stream and the isolated agent audio) under one configuration, one judge set and one bootstrap procedure. Conversational dynamics. Turn-taking (TT) scores every floor transfer’s offset on an on-time curve chosen by the transfer’s kind, and a single unanswered user turn zeroes the conversation; it is ported verbatim from EVA-Bench (Bogavelli et al., 2026) and is the one metric in the suite where silence is penalised rather than rewarded. Conversation progression (CPCP) is one topic-free judge pass over the transcript. Selectivity (SELSEL) is the fraction of injected distractor events correctly ignored, scored against gold labels written at injection time. All three read only the transcript and timing log, so identical code runs in every world. Agentic capability. The goal-state check GSGS asks whether the world ended in the right state: a canonical hash of the record store against a gold replay, and in Pathfinding the walker’s final position against the destination. The reward multiplies exactly the binary factors each scenario names (GSGS, a gold-action match ACTACT, and, where what is said or refused is the point, a judged assertion NLANLA); in Pathfinding, which has no gold action list, an efficiency conjunct (η≥0.75η≥ 0.75) replaces them so that arrival alone is not enough. Pass@1Pass@1 is the mean single-run reward, Pass@3Pass@3 the at-least-once reward over three draws, and Pass3Pass^3 τ-bench’s PasskPass^k at k=3k=3 (Yao et al., 2024); Appendix H gives the estimators. Effort is measured against each task’s own reference workload: the over-effort ratio ϱ+ ^+ and the under-effort share π−π^-, each reported with its conditioning share. A system that never acts scores near-perfect selectivity and competitive turn-taking, so experience metrics are read jointly with π−π^-; a null reference policy ships with the release as a constructive check. Naturalness. Faithfulness (FAIFAI) is one judge pass over the transcript with the agent’s instructions, role and tool schemas, five binary dimensions scored as their minimum. DNSMOS (Reddy et al., 2022), UTMOS (Saeki et al., 2022) and NISQA (Mittag et al., 2021) are no-reference MOS predictors over the isolated agent channel. 5 Results and Discussion Table 3 carries the full suite less the effort pair; Table 4 carries the effort pair. We read the results in three passes: what the main table says about each capability pillar and how the pillars relate (Section 5.1), how the systems fail (Section 5.2), and what the auxiliary instruments separate that headline metrics cannot (Section 5.3). The remaining analyses (the deep dives, the range analysis, the metric–reward associations, and the harness dials) are in Appendices I–L. Agentic Capability Conversational Dynamics Naturalness GSGS Pass@1Pass@1 Pass@3Pass@3 Pass3Pass^3 TT CPCP SELSEL FAIFAI M^D M_D M^U M_U M^N M_N 2S_2S 0.415±.081 0.037±.033 0.037±.056 0.037±.056 0.760±.043 1.022±.030 0.976±.014 1.911±.133 3.152±.043 2.565±.054 2.550±.047 3.1_3.1 0.585±.081 0.207±.067 0.363±.161 0.085±.094 0.449±.033 1.437±.096 0.759±.053 1.585±.104 3.402±.012 3.398±.018 3.522±.023 2.1_2.1 0.689±.078 0.200±.070 0.352±.167 0.063±.072 0.763±.028 1.556±.100 0.340±.051 2.037±.122 3.360±.015 4.081±.049 3.589±.033 mini_mini 0.385±.081 0.126±.056 0.233±.141 0.044±.061 0.535±.051 1.200±.078 0.470±.046 1.548±.111 3.198±.050 3.530±.178 3.287±.115 VTF_VTF 0.726±.074 0.326±.081 0.511±.169 0.144±.104 0.686±.033 1.644±.119 0.613±.057 1.800±.126 3.143±.015 3.657±.016 3.117±.020 2S_2S 0.052±.037 0.000±.000 0.000±.000 0.000±.000 0.631±.048 1.030±.026 0.968±.018 1.911±.141 3.145±.037 2.587±.038 2.629±.045 3.1_3.1 0.630±.081 0.489±.085 0.715±.150 0.270±.150 0.443±.037 1.874±.119 0.727±.062 2.015±.133 3.373±.014 3.423±.017 3.484±.023 2.1_2.1 0.637±.081 0.533±.081 0.704±.150 0.381±.170 0.741±.036 1.867±.115 0.369±.057 2.252±.126 3.361±.007 4.112±.007 3.601±.014 mini_mini 0.333±.081 0.296±.078 0.474±.159 0.152±.120 0.702±.049 1.681±.122 0.350±.058 1.763±.137 3.387±.007 4.163±.008 3.727±.014 VTF_VTF 0.652±.081 0.519±.081 0.685±.159 0.341±.154 0.680±.030 2.015±.126 0.627±.062 2.148±.126 3.132±.013 3.692±.017 3.109±.020 2S_2S 0.385±.081 0.007±.011 0.022±.033 0.000±.000 0.611±.055 1.007±.011 0.973±.020 1.437±.100 3.202±.016 2.634±.031 2.643±.035 3.1_3.1 0.652±.078 0.378±.081 0.674±.148 0.119±.102 0.336±.031 1.378±.085 0.742±.053 1.667±.115 3.370±.023 3.372±.045 3.459±.038 2.1_2.1 0.615±.081 0.511±.085 0.789±.119 0.244±.133 0.649±.034 1.585±.122 0.391±.053 2.015±.133 3.353±.007 4.082±.009 3.581±.015 mini_mini 0.407±.085 0.074±.044 0.137±.111 0.037±.056 0.494±.050 1.215±.085 0.498±.059 1.652±.126 3.349±.013 4.067±.048 3.642±.035 VTF_VTF 0.793±.067 0.519±.085 0.715±.154 0.304±.150 0.612±.027 1.948±.119 0.704±.054 1.859±.133 3.128±.015 3.647±.050 3.040±.027 2S_2S 0.385±.081 0.015±.019 0.033±.050 0.000±.000 0.637±.052 1.007±.011 0.960±.022 1.556±.115 3.191±.017 2.655±.029 2.671±.040 3.1_3.1 0.630±.081 0.363±.085 0.656±.152 0.089±.076 0.249±.023 1.422±.096 0.707±.054 1.474±.107 3.398±.010 3.425±.016 3.472±.021 2.1_2.1 0.556±.081 0.481±.081 0.711±.159 0.211±.120 0.551±.035 1.296±.078 0.430±.051 1.881±.126 3.336±.006 4.087±.006 3.569±.012 mini_mini 0.407±.081 0.244±.074 0.530±.141 0.030±.030 0.465±.047 1.185±.078 0.476±.050 1.519±.115 3.363±.006 4.128±.007 3.658±.014 VTF_VTF 0.807±.067 0.556±.081 0.811±.119 0.278±.130 0.553±.019 1.689±.126 0.680±.057 1.778±.126 3.090±.014 3.665±.014 3.012±.017 2S_2S 0.341±.081 0.007±.011 0.022±.033 0.000±.000 0.695±.054 1.022±.030 0.995±.007 1.748±.119 3.175±.027 2.444±.038 2.385±.039 3.1_3.1 0.881±.056 0.459±.081 0.704±.152 0.204±.119 0.314±.025 1.778±.111 0.735±.058 2.067±.111 3.366±.025 3.372±.017 3.472±.030 2.1_2.1 0.815±.067 0.674±.081 0.930±.054 0.374±.143 0.656±.022 1.667±.111 0.404±.056 2.319±.122 3.350±.006 4.085±.007 3.601±.013 mini_mini 0.585±.081 0.296±.078 0.541±.157 0.074±.052 0.541±.043 1.407±.096 0.488±.056 1.970±.122 3.372±.006 4.134±.007 3.685±.012 VTF_VTF 0.830±.063 0.489±.081 0.652±.165 0.307±.152 0.597±.017 2.207±.111 0.647±.056 2.207±.122 3.136±.023 3.709±.015 3.086±.022 2S_2S 0.000±.000 0.000±.000 0.000±.000 0.000±.000 0.062±.058 1.022±.033 0.991±.007 1.244±.167 3.16±.061 2.45±.052 2.49±.031 3.1_3.1 0.978±.033 0.489±.133 0.722±.261 0.222±.200 0.538±.029 1.178±.111 0.782±.067 1.511±.167 3.36±.021 3.42±.022 3.45±.031 2.1_2.1 0.400±.144 0.200±.111 0.467±.283 0.000±.000 0.561±.084 1.289±.156 0.551±.078 1.644±.189 3.34±.016 4.12±.026 3.66±.045 mini_mini 0.311±.133 0.089±.078 0.233±.200 0.000±.000 0.387±.102 1.156±.122 0.526±.068 1.333±.189 3.33±.020 4.11±.028 3.67±.041 VTF_VTF 0.867±.100 0.533±.133 0.789±.200 0.222±.133 0.682±.065 1.378±.133 0.507±.068 1.489±.156 3.13±.013 3.75±.016 3.20±.016 all six 2S_2S 0.263±.028 0.011±.007 0.019±.019 0.006±.009 0.566±.021 1.019±.012 0.977±.006 1.635±.054 3.172±.015 2.556±.017 2.562±.017 3.1_3.1 0.726±.029 0.398±.039 0.639±.073 0.165±.054 0.388±.012 1.511±.043 0.742±.024 1.720±.051 3.378±.008 3.402±.010 3.477±.012 2.1_2.1 0.619±.037 0.433±.035 0.659±.068 0.212±.048 0.653±.018 1.543±.048 0.414±.024 2.025±.057 3.350±.004 4.095±.010 3.600±.010 mini_mini 0.405±.038 0.188±.028 0.358±.065 0.056±.028 0.521±.025 1.307±.041 0.468±.023 1.631±.054 3.334±.009 4.022±.033 3.611±.022 VTF_VTF 0.779±.031 0.490±.039 0.694±.068 0.266±.057 0.635±.015 1.814±.051 0.630±.024 1.880±.052 3.127±.007 3.687±.010 3.093±.009 Table 3: Main results across the six worlds (realistic channel; n=135n=135 conversations per enterprise cell, 45 per Pathfinding cell). Worlds by mark ( Banking, Logistics, Healthcare, Insurance, Travel, Pathfinding); systems by vendor logo and version. Subscripts are 95% bootstrap half-widths. Bold marks the best system per world and column. GSGS is the goal-state check; Pass@1Pass@1, Pass@3Pass@3 and Pass3Pass^3 are the single-run reward, at least one pass in three draws, and all three of three. CPCP and FAIFAI are on 1–3, MOS predictors on 1–5; the all six block is the unweighted mean over worlds. Effort metrics are in Table 4. 5.1 Three capabilities, one table Table 3 asks three questions of the same five systems: did the task get done (Agentic Capability), how did the conversation go (Conversational Dynamics), and how did it sound (Naturalness). To our knowledge this is the first evaluation to score all three side by side for full-duplex voice agents, so the table is also a first answer to whether they travel together. On the agentic pillar the all six rows order the systems cleanly: Grok Voice Think Fast 1.0 leads (Pass@1Pass@1 0.490, GSGS 0.779), GPT-Realtime-2.1 follows (0.433), Gemini-3.1-Flash-Live is third (0.398), GPT-Realtime-2.1-mini completes less than half of what its full-size sibling does (0.188), and Nova 2 Sonic barely registers (0.011). The level is a property of the world as much as of the system: GPT-Realtime-2.1 alone spans 0.200 in Banking to 0.674 in Travel, worlds that differ only in subject matter, while the ordering barely moves (Appendix J). Reliability decays fast: Pass3Pass^3 is roughly half of Pass@1Pass@1 for the two leaders and under a third for GPT-Realtime-2.1-mini, and Figure 1 shows the same decay at every k; no system passes even one scenario in five reliably. Conversational Dynamics does not reproduce that order. GPT-Realtime-2.1 holds the best turn-taking (0.653 pooled), Grok Voice Think Fast 1.0 the best conversation progression (1.814), and Gemini-3.1-Flash-Live, third on the reward, is last on turn-taking (0.388) because it pauses to call tools. Nova 2 Sonic is the sharpest dissociation in the table: respectable turn-taking (0.566) and the best selectivity in every world (0.960–0.995) sit beside an agentic column of near-zeros. Being good in the conversation and being good at the task are different capabilities, a system can hold the floor gracefully while doing nothing, and a leaderboard built on dynamics alone would rank the least capable system near the top. Naturalness separates least. All five systems sit within a quarter point on DNSMOS (3.13–3.40), and the two MOS leaders, GPT-Realtime-2.1 and GPT-Realtime-2.1-mini, differ by 2.3×2.3× on the reward: how a system sounds carries almost no information about what it completes. We read this two ways. Production systems have largely converged on acoustic quality, so this pillar now needs sharper instruments, expressiveness, empathy and prosodic appropriateness rather than signal quality. And a deployment decision made on perceived quality alone will pick the wrong system; faithfulness (FAIFAI), judged from content rather than sound, is the one naturalness column that still tracks the reward. 5.2 Failure modes across the worlds ϱ+(π+) ^+\,(π^+) ϱ−(π−) ^-\,(π^-) ϱ+(π+) ^+\,(π^+) ϱ−(π−) ^-\,(π^-) 2S_2S 1.778±.5832†1.778_±.583^2 0.112±.037950.112_±.037^95 2S_2S 2.061±.349272.061_±.349^27 0.146±.057580.146_±.057^58 3.1_3.1 2.126±.200612.126_±.200^61 0.507±.072290.507_±.072^29 3.1_3.1 1.791±.176671.791_±.176^67 0.547±.100170.547_±.100^17 2.1_2.1 1.804±.158531.804_±.158^53 0.317±.104260.317_±.104^26 2.1_2.1 1.860±.145831.860_±.145^83 0.365±.2146†0.365_±.214^6 mini_mini 1.840±.149531.840_±.149^53 0.393±.078410.393_±.078^41 mini_mini 2.223±.203762.223_±.203^76 0.458±.1757†0.458_±.175^7 VTF_VTF 1.841±.160561.841_±.160^56 0.498±.128210.498_±.128^21 VTF_VTF 1.895±.174781.895_±.174^78 0.375±.2364†0.375_±.236^4 2S_2S 1.600±.11711†1.600_±.117^11 0.189±.050770.189_±.050^77 2S_2S 2.322±.439162.322_±.439^16 0.161±.050640.161_±.050^64 3.1_3.1 1.791±.147601.791_±.147^60 0.655±.07710†0.655_±.077^10 3.1_3.1 2.748±.267782.748_±.267^78 0.467±.1753†0.467_±.175^3 2.1_2.1 1.728±.136511.728_±.136^51 0.644±.1197†0.644_±.119^7 2.1_2.1 2.615±.220762.615_±.220^76 0.315±.2157†0.315_±.215^7 mini_mini 1.789±.150501.789_±.150^50 0.603±.071160.603_±.071^16 mini_mini 3.002±.361713.002_±.361^71 0.570±.12713†0.570_±.127^13 VTF_VTF 1.828±.161521.828_±.161^52 0.652±.0958†0.652_±.095^8 VTF_VTF 2.764±.262762.764_±.262^76 0.583±.2174†0.583_±.217^4 2S_2S 1.907±.36910†1.907_±.369^10 0.212±.053780.212_±.053^78 2S_2S 4.250±3.71618†4.250_± 3.716^18 0.439±.060730.439_±.060^73 3.1_3.1 2.273±.217732.273_±.217^73 0.411±.1729†0.411_±.172^9 3.1_3.1 2.169±.279802.169_±.279^80 0.595±.11913†0.595_±.119^13 2.1_2.1 1.985±.165731.985_±.165^73 0.626±.1065†0.626_±.106^5 2.1_2.1 1.476±.25440†1.476_±.254^40 0.637±.058470.637_±.058^47 mini_mini 2.286±.266592.286_±.266^59 0.511±.095190.511_±.095^19 mini_mini 1.524±.23142†1.524_±.231^42 0.681±.065510.681_±.065^51 VTF_VTF 1.899±.141671.899_±.141^67 0.479±.2056†0.479_±.205^6 VTF_VTF 1.640±.165471.640_±.165^47 0.672±.077440.672_±.077^44 Table 4: Effort against the task’s reference workload (ideal tool-call count in the enterprise worlds, optimal block count in Pathfinding): over-effort ϱ+ ^+ and under-effort ϱ− ^- ratios, conditioning shares π±π^± (%) as superscripts; † marks fewer than 20 conversations. Section 5.2 reads the table; 2S_2S’s ϱ+=4.25 ^+=4.25 rests on 8 conversations. Table 4 measures every conversation against the task’s own reference workload, the ideal tool-call count in the enterprise worlds and the optimal block count in Pathfinding. In the enterprise worlds the engaged systems fail by doing too much: they exceed the ideal workload in 50–83% of conversations, with over-effort ratios of 1.7–3.0, largest in Travel, the world whose modal right answer is to do nothing. The disengaged system fails the opposite way, falling short of the workload in 58–95% of its enterprise conversations. Pathfinding inverts the pattern. There the strongest systems over-work ( Gemini-3.1-Flash-Live exceeds its budget in 80% of conversations and still arrives most often) while the weak systems under-work, because a copilot that has lost the walker stops issuing instructions. Across all worlds the under-effort share π−π^- is the strongest process predictor of failure in the suite (Appendix K); it cannot be won by silence, because it is signed against the workload, and we would recommend the field report it beside every completion score. 5.3 What the auxiliary metrics reveal In the enterprise worlds, the clearest auxiliary signal is the credential word error rate (Table 13, appendix). GPT-Realtime-2.1 hears best: 10.4% error on spoken credentials over the telephone channel, against 14.4 for Grok Voice Think Fast 1.0, 15.5 for Gemini-3.1-Flash-Live and 20.2 for GPT-Realtime-2.1-mini, so distillation roughly doubles credential error inside one family; Nova 2 Sonic’s 33.3% is measured only on the minority of calls in which it attempted a credential at all. Pooled, about one spoken credential in six is mis-heard (17.1%), which puts a floor under identity verification for every system: recognition over a telephony channel, not reasoning, is still a first-order bottleneck. Pathfinding adds an instrument no enterprise world can: the exploration share, the fraction of walker moves made under uncertainty rather than on a confident instruction (Table 14, appendix). The world is, in effect, an MDP, junctions for states and maneuvers for actions, so we read each system’s behaviour as a policy through the explore–exploit lens of the agents literature. Systems probe with sign-readings, look-arounds and trial walks, then commit, and the policies differ sharply by system and far less by condition. Gemini-3.1-Flash-Live and Grok Voice Think Fast 1.0 explore on 16.2% and 22.7% of moves; GPT-Realtime-2.1-mini, GPT-Realtime-2.1 and Nova 2 Sonic on 40.3, 43.4 and 51.2; realistic acoustics raise exploration for most systems ( GPT-Realtime-2.1 from 41.1 to 50.0 on single intent, Grok Voice Think Fast 1.0 from 12.4 to 24.9 on rerouting) without reordering them, and exploration falls for most systems on all-day assistance, where the second leg retraces known ground. Exploration is not the virtue here that it is in reinforcement learning; in this corpus the trade-off resolves entirely toward exploitation. In every condition and channel the two lowest explorers are the two most likely to arrive, the highest explorer never arrives at all, and only six non-arrivals in the whole realistic corpus end at a wrong destination; the rest run out of clock (Appendix E.1). Exploration in Pathfinding is mostly rambling: motion spent recovering a belief about the walker that the better systems never lost, and Pass@1Pass@1 beside the exploration column is what says the wandering policies do not pay. 6 Conclusion and Future Work We introduced DuplexWorld, a benchmark that evaluates S2S voice agents across six worlds and eleven conversation types under one harness and one twelve-metric suite, and evaluated five commercial systems over 3,825 conversations. The results are direct: even the best systems leave substantial room on every axis, with no system exceeding Pass@1Pass@1 0.490, turn-taking 0.653 or DNSMOS 3.378 pooled over the worlds (Figures 7 and 9). The analysis yields four concrete findings. A score depends on the world and the conversation type as much as on the system: one system spans Pass@1Pass@1 0.200 to 0.674 across worlds that differ only in subject matter. Conversational and agentic capability are separate: the best conversationalists are not the best task-completers, and a system can hold the floor gracefully while completing nothing (Figure 8). Acoustic quality tells you almost nothing about competence, while mis-heard credentials put a floor under identity verification for every system. And in Pathfinding, read through an explore–exploit lens, the systems that explore most arrive least (Figure 10). The practical takeaway for deployment is equally direct: report the under-effort share and the harness configuration beside every score, because both move the numbers. Future work includes a human-rating study to validate the judge-based metrics, campaigning the released d_vague variant (Appendix E), a factorial study over the three harness dials (Appendix I), and extending the benchmark to more languages and real recorded speech. Limitations Pathfinding is a single world that raises several demands at once (who holds the effectors, what the agent can observe, and whether the world moves between utterances), so no single demand is separately identified, and the five-world reference interval bounds the subject-matter explanation without eliminating it. Its headline cells rest on 45 conversations per system against 135 per enterprise cell, its reward carries an efficiency threshold whose sensitivity Table 9 bounds, and 215 of its 450 conversations end at the step cap, with one provider’s session cap censoring that system’s cells (Appendix E.1); the strongest cross-world claims are therefore restricted to the four metrics computed by identical code everywhere. Caller speech is synthesised rather than recorded, which weakens conclusions about accents and disfluency; the available real-speech corpora (Lin et al., 2026a; Zhao et al., 2026) do not provide goal-directed tasks with a verifiable end state, a limitation τ-Voice and EVA-Bench share (Ray et al., 2026; Bogavelli et al., 2026). The benchmark is English only, simulated users are more patient than real ones, tools are declarative mocks with zero latency, and the reward is binary with no partial credit. Our judge-based metrics are assembled from instruments whose authors validated them against human ratings in their source settings (Bogavelli et al., 2026; Reddy et al., 2022; Saeki et al., 2022; Mittag et al., 2021); we have not re-validated them on this corpus, so the associations we report are between our metrics rather than between a metric and a human, and GSGS is passable by inaction wherever the gold terminal state equals the seeded state. An earlier version of the benchmark included a seventh world; we cut it during task-design iteration and record the cut rather than let a reader assume the design space is covered by these six. Ethics Statement Potential risks. A benchmark that ranks commercial voice agents can be used to justify deploying one on consequential work. Two of our results argue directly against reading it that way: metrics a system can win by declining to act do not track capability, and the best-sounding systems in this study are not the most capable ones (Section 5). Perceived quality is not evidence of competence here, and a procurement decision that uses one as a proxy for the other will select the wrong system. We would rather this finding travel than the leaderboard. No human subjects. Every caller utterance in DuplexWorld is synthesised; no human speech was recorded, solicited or replayed at any point, so no consent, compensation or institutional-review question arises. We state this explicitly because the benchmark is about telephone conversations and a reader is entitled to assume otherwise. The judge- and predictor-based instruments we use were validated against human ratings by their authors in their source settings (Bogavelli et al., 2026; Reddy et al., 2022; Saeki et al., 2022; Mittag et al., 2021). Synthetic personas are not a speaker population. The personas carry accent, gender and speaking-rate attributes, and systems do not perform equally across them; Appendix I reports personas some configurations could not hear at all. Per-persona results ship with the release so this is visible, but the persona set represents no real population, and a per-persona gap in DuplexWorld is evidence about DuplexWorld’s personas rather than about the speakers they resemble. All task data is fabricated. Records, policies, identifiers and transcripts are authored; no customer data or personal information of any kind enters the corpus, and the release pass strips any authored value that collides with a real institution, address or number space. Cost and energy. Every number in this paper comes from paid inference against commercial realtime APIs; the corpus totals 387 hours of simulated speech (Appendix P), so the scale of any reproduction is known before it is incurred. References S. Arora, Z. Lu, C. Chiu, R. Pang, and S. Watanabe (2025) Talking turns: benchmarking audio foundation models on turn-taking dynamics. In International Conference on Learning Representations (ICLR), External Links: Link Cited by: §2, Table 1. V. Barres, H. Dong, S. Ray, X. Si, and K. Narasimhan (2025) τ2τ^2-Bench: evaluating conversational agents in a dual-control environment. External Links: 2506.07982, Link Cited by: Appendix B, §2, Table 1. T. Bogavelli, G. Gauthier Melançon, K. Stankiewicz, O. Bamgbose, F. Riols, H. H. Nguyen, R. Mehndiratta, L. D. Brin, J. Marinier, H. Subramani, A. Madamala, S. K. Nemala, and S. Sunkara (2026) EVA-Bench: a new end-to-end framework for evaluating voice agents. External Links: 2605.13841, Link Cited by: Table 5, Appendix R, Appendix B, Appendix H, Appendix I, §1, §2, Table 1, §3, §4.2, §4.3, Limitations, No human subjects.. C. E. Jimenez, J. Yang, A. Wettig, S. Yao, K. Pei, O. Press, and K. Narasimhan (2024) SWE-bench: can language models resolve real-world GitHub issues?. In International Conference on Learning Representations (ICLR), Cited by: Table 5. P. Liang, R. Bommasani, T. Lee, et al. (2023) Holistic evaluation of language models. Transactions on Machine Learning Research. Cited by: Table 5. G. Lin, C. Chen, Z. Chen, and H. Lee (2026a) Full-duplex-bench-v3: benchmarking tool use for full-duplex voice agents under real-world disfluency. External Links: 2604.04847, Link Cited by: Appendix D, §1, §2, Table 1, Limitations. G. Lin, S. S. Kuan, J. Shi, K. Chang, S. Arora, S. Watanabe, and H. Lee (2025a) Full-duplex-bench-v2: a multi-turn evaluation framework for duplex dialogue systems with an automated examiner. External Links: 2510.07838, Link Cited by: §1, §2. G. Lin, S. S. Kuan, Q. Wang, J. Lian, T. Li, S. Watanabe, and H. Lee (2026b) Full-duplex-bench v1.5: evaluating overlap handling for full-duplex speech models. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), External Links: Link Cited by: §1, §2. G. Lin, J. Lian, T. Li, Q. Wang, G. Anumanchipalli, A. H. Liu, and H. Lee (2025b) Full-duplex-bench: a benchmark to evaluate full-duplex spoken dialogue models on turn-taking capabilities. External Links: 2503.04721, Link Cited by: §1, §2, Table 1. J. Lu, Y. Wang, J. Luo, Y. Chen, T. Liang, S. Ji, Z. Jiang, X. Yang, Y. Zhang, X. Cheng, C. Wen, C. Pan, H. Wang, C. Ye, J. Wu, X. Jiang, G. Jiang, and Z. Zhao (2026) A survey of full-duplex spoken dialogue systems: architectural hierarchy, interaction ontology, and decision state machine. External Links: 2606.19453, Link Cited by: §1, §2. G. Mittag, B. Naderi, A. Chehadi, and S. Möller (2021) NISQA: a deep CNN-self-attention model for multidimensional speech quality prediction with crowdsourced datasets. In Proceedings of Interspeech, Cited by: §4.3, Limitations, No human subjects.. S. N. Modi, G. Mahajan, M. Wetter, and R. Welles (2026) EchoChain: a full-duplex benchmark for state-update reasoning under interruptions. External Links: 2604.16456, Link Cited by: Appendix D, §1, §2, Table 1. S. Ray, K. Dhandhania, V. Barres, and K. Narasimhan (2026) τ-Voice: benchmarking full-duplex voice agents on real-world domains. External Links: 2603.13686, Link Cited by: Appendix B, §1, §2, Table 1, §4.2, Limitations. C. K. A. Reddy, V. Gopal, and R. Cutler (2022) DNSMOS P.835: a non-intrusive perceptual objective speech quality metric to evaluate noise suppressors. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Cited by: §4.3, Limitations, No human subjects.. T. Saeki, D. Xin, W. Nakata, T. Koriyama, S. Takamichi, and H. Saruwatari (2022) UTMOS: UTokyo-SaruLab system for VoiceMOS challenge 2022. In Proceedings of Interspeech, Cited by: §4.3, Limitations, No human subjects.. A. Salimi, W. Ma, Y. Tang, D. Shen, M. Li, and A. Smola (2026) IHBench: evaluating post-interruption recovery in voice agents with structured workflows. External Links: 2606.19595, Link Cited by: Appendix D, §1, §2, Table 1. C. Wang, H. Xue, G. Li, Z. Zhao, S. Wang, S. Wang, X. Xu, H. Bu, and L. Xie (2026) Full-duplex interaction in spoken dialogue systems: a comprehensive study from the ICASSP 2026 HumDial challenge. External Links: 2604.21406, Link Cited by: §2. S. Yao, N. Shinn, P. Razavi, and K. Narasimhan (2024) τ-Bench: a benchmark for tool-agent-user interaction in real-world domains. External Links: 2406.12045, Link Cited by: Appendix H, §2, Table 1, §3, §4.3. Z. Zhao, S. Wang, G. Li, H. Xue, C. Wang, S. Wang, L. Xiao, Z. Zhang, H. Bu, X. Xu, X. Wang, H. Liu, E. S. Chng, H. Lee, H. Li, and L. Xie (2026) The ICASSP 2026 HumDial challenge: benchmarking human-like spoken dialogue systems in the LLM era. External Links: 2601.05564, Link Cited by: §2, Table 1, Limitations. Appendix A Nomenclature term what it denotes # prior name world one enterprise support line, or the pedestrian environment; its own tools, policy corpus and record schema six domain; environment type a conversation shape; a taxonomy of kinds, deliberately not a difficulty ladder. Types 1–9 are shared by the five enterprise worlds; rerouting and all-day assistance arise only in Pathfinding (release names blocked, errand; single intent there is base) eleven task category; condition scenario user goal, persona, seeded world state and gold terminal state, jointly consistent 156 scenario; task instance channel the acoustic path a scenario is run over: clean or realistic 2 perturbation run one execution of one scenario by one system over one channel 5 / scen. trial pillar a group of metrics reported together and never composed into a scalar 3 −- Table 5: One word per level. The last column is the nearest term a reader is likely to arrive from, in the sense of EVA-Bench (Bogavelli et al., 2026), SWE-bench (Jimenez et al., 2024) and HELM (Liang et al., 2023). Appendix B Harness implementation Our harness follows the tick-based design established by τ-Voice and τ2τ^2-bench (Ray et al., 2026; Barres et al., 2025), and we build directly on their released orchestrator rather than reimplementing it. Simulated time advances in 200 ms ticks, so model latency never shifts an event’s position in the timeline and a slow model and a fast model produce comparable traces; each side hears the previous tick, and tool results deliver on the next tick so audio never stalls on tool latency. Task and evaluation scaffolding, the record-store abstraction, goal-state grading and PasskPass^k are inherited from τ2τ^2-bench; from EVA-Bench we port the turn-taking metric verbatim, including its unanswered-turn rule, and adopt its practice of measuring user-simulator goal adherence without letting it gate reward (Bogavelli et al., 2026). Building on a shared harness is deliberate: it holds the measurement apparatus fixed while the questions change. Our own additions are confined to five places: a silence and suspension sentinel that makes dead air scorable in the suspension type; bipartite gold-action matching (presence only, so an omission fails while surplus does not) for policy refusal; distractor injection with gold labels written at injection time, which is what makes selectivity scorable against ground truth; a voice preflight that screens every persona against the server VAD threshold before any scored run, which Section I shows is not optional; and an annotation layer over the tick stream that attributes anomalous runs to agent, simulator or harness before scoring. Of the eleven conversation types only two require any of these to be scorable at all, and the two Pathfinding types needed a new world rather than a harness change. The harness is infrastructure for this work rather than a contribution of it. Infrastructure failure means rerun and never score; agent non-response is a valid scored end state, and agent freeze is a hard zero on TT through the unanswered-turn term. Credentials are matched on their spoken renderings, so “three four five” and “345” are one value. One provider-side artefact is handled at this layer: Nova 2 Sonic’s Pathfinding sessions are terminated by the provider’s own session cap (Appendix E.1). role model notes user simulator gpt-5.6-luna (Azure) writes the caller’s or walker’s words; provider-default sampling decision model claude-haiku-4.5 (OpenRouter) “should I interrupt” and “should I backchannel” as two independent binary calls every ∼2.0 2.0 s while the agent speaks; realistic channel only judges per metric Table 12, Appendix H Table 6: Harness-side models, uniform across every world and every reported number. Turn-taking control sits entirely on the user side; the agent under test is never told any threshold. Appendix C Systems and configuration All five systems are served over WebSocket realtime endpoints through one adapter. llm_args is null on every scored run: no generation parameters are set anywhere, so every system runs at provider defaults. No vendor returns a dated snapshot identifier through the client, so the strings of Table 2 are the family identifiers the APIs report. Pathfinding access windows, first-to-last scored run: gpt-realtime-2.1 2026-07-30 to 08-03, gpt-realtime-2.1-mini 2026-07-31 to 08-03, gemini-3.1-flash-live-preview 2026-08-01 to 08-03, xai-realtime 2026-08-01 to 08-03, amazon.nova-2-sonic-v1:0 2026-08-02. The matched configuration across all six worlds: speech complexity regular, step cap 1200 s, wall-clock timeout 1200 s, server VAD threshold 0.2, fixed seed, persona pinned per scenario (Appendix G). Appendix D World specifications: enterprise worlds world governing regime authority granularity refusal character what it isolates Banking financial-crime statute per caller, stepped up by operation risk tier statutory Translating a symptom into an operation when several plausible, sympathetically framed requests are reporting-threshold tripwires. Logistics contract topology over physical goods per the shipment’s contractual parties contractual, physical Deciding who is entitled to redirect goods the agent can move in the record but cannot see. Healthcare lawful disclosure per record, via a disclosure matrix privacy Refusing without leaking in the act of refusing. A fully verified caller may be entitled to nothing. Insurance institutional rules forbidding adjudication per policyholder institutional Taking first notice of a loss the agent is structurally forbidden from deciding. Travel the platform’s own limited remit between two parties per delegate list and co-host roster scope Doing nothing, correctly: the modal right answer is a factual correction or a refusal held with nowhere to escalate to. Pathfinding perceptual grounding −- −- Grounding an instruction against a heading neither party can state. Table 7: The six regimes: what governs each world, at what granularity authority is decided, what refusal looks like there, and the capability each isolates. type the shape it fixes, and what it tests single intent One request, one resolution path. The control condition; in Pathfinding, a base route of five blocks and two turns. multi intent Several requests, resolvable in more than one order, with at least one dependency: goal decomposition under a live conversation. policy refusal A well-formed request that must be refused. Correctness is a property of what is withheld. record disambiguation More than one record matches what the caller said; acting before disambiguating is the failure. identity verification Authority must be established before anything is read or written, and a failed check must not leak the record’s existence. guided procedure The effectors move to the user: the agent instructs, waits, and must trust an unverifiable human report as the trigger for its own write. mid-call correction The caller revises a premise after the agent has begun acting on it. suspension The call is interrupted and resumed across a gap the agent must hold state over; the one type in which dead air is itself scored. narrative intake A long unstructured account from which the agent must extract, confirm and file the structured facts. rerouting Closures missing from the map seal every visible route; querying early returns nothing, so a belief formed from a tool result expires silently. Arises only in Pathfinding. all-day assistance A second goal emerges on arrival at the first; the test is staying with the user, and useful, over a long horizon. Arises only in Pathfinding. Table 8: The eleven conversation types. Types 1–9 are shared by the five enterprise worlds; the last two arise only in Pathfinding, whose episodes are all structurally guided procedure with one of single intent, rerouting or all-day assistance layered on top (release names base, blocked, errand; Appendix D). Each enterprise world is one instantiation of the telephone-support construct of Section 3.1; Table 7 gives the per-world regimes and Table 8 the conversation types. The unsafe behaviour in several regimes is unrepresentable rather than discouraged, because it is enforced in the tool surface: Banking contains no tool that would require an employee to hear a one-time code aloud, Healthcare gates claim reads on the verified member so that authority is decided per record, and Travel scores restraint through negative assertions and write-counting rather than end-state checks. Each scenario names its reward basis B(x)⊆GS,ACT,NLAB(x) \GS,ACT,NLA\, and the per-scenario bases, together with the difficulty-governing parameters that live inside a type (escalation count, candidate-record count, interruption count), ship with the release. All six worlds share one volatility stressor: the world can change with no notification, discoverably only by re-querying. This exposes three sources of divergence between what the agent believes and what is true, and each world instantiates all three: the world changes (a record moves mid-call; a closure appears on the planned route), the user’s intent changes (a premise is revised; a second destination surfaces), or the user’s execution diverges (the caller misacts on an instruction; the walker takes the wrong turn). The three demand different repairs, and prior interruption work varies only the second (Lin et al., 2026a; Modi et al., 2026; Salimi et al., 2026). Appendix E World specifications: Pathfinding World. An 8×88× 8 rectilinear grid: 64 junctions and 105 walkable segments, with two superblocks removing 7 edges so that naive rectilinear reasoning fails. Block spacing is uniform (274 m avenue, 81 m street) and time advances in 200 ms ticks to a 6,000-tick cap, twenty simulated minutes. The paper’s type names map to the release as single intent == base (five blocks, two turns, no closures), rerouting == blocked (four closures absent from the copilot’s map seal every visible route; exactly one seven-block route survives), and all-day assistance == errand (two legs of 5+65+6 blocks, the second stop revealed only on arrival at the first). Tool surface. Walker motion is maneuver-driven: one maneuver call moves one street corner, and the walker keeps talking while moving. Nine walker-side maneuvers (turns, walking, crossing, sign-reading, looking around) are paired with five copilot-side reads (GPS, place lookup, directions, live status, and a write-only belief probe). All returns are natural-language strings rather than structured schemas: GPS is described as accurate to about 40 metres and never includes heading, and live status reports closures only within two blocks of the walker, so querying early returns nothing and a belief formed from a tool result expires silently. The heading constraint. Heading is ground truth in the environment and never exposed: no walker tool returns it, and the directions call takes no facing argument and returns absolute bearings. The walker’s refusal of compass words is enforced by prompt only, so the constraint is structural on the tool surface and prompt-level on the persona; in 68% of rerouting calls the copilot still asks a variant of “which way is north”, and the walker’s refusal is what holds. Arrival scoring. Scoring is a single post-hoc environment assertion, deliberately not exposed as a callable tool. It requires the walker to be standing at the destination node when the call ends rather than merely to have passed through it; latching on first touch was rejected after a probe conversation scored 1.0 for a walker led eight blocks past the goal. Pass@1Pass@1 at efficiency threshold η0 _0 system GSGS 0.500.50 0.600.60 0.750.75 0.900.90 1.001.00 2S_2S 0.000 0.000 0.000 0.000 0.000 0.000 3.1_3.1 0.978 0.800 0.667 0.489 0.422 0.422 2.1_2.1 0.400 0.378 0.356 0.200 0.156 0.156 mini_mini 0.311 0.200 0.156 0.089 0.044 0.044 VTF_VTF 0.867 0.756 0.667 0.533 0.378 0.378 all 0.511 0.427 0.369 0.262 0.200 0.200 Table 9: Sensitivity of Pass@1Pass@1 to the efficiency threshold η0 _0, Pathfinding, realistic channel. A run passes if the walker arrived and η≥η0η≥ _0; GSGS is arrival alone. Bold is the operating point used throughout (η0=0.75 _0=0.75). The ordering is unchanged at every threshold except η0=0.75 _0=0.75, where VTF_VTF and 3.1_3.1 swap on a 0.044 margin, so no conclusion rests on the choice: it sets the level, not the order. The weak systems are not failing on efficiency; 95% of their non-arrivals are step-cap endings. Exploration share. The exploration share of Table 14 is the fraction of walker moves made under uncertainty rather than on a confident instruction, labelled per move by gpt-5.6-luna over the tick stream; all 30 condition–channel cells are labelled. The replay also computes a heading-localisation trace per run: position is essentially solved by every system (accuracy 0.93–1.00) while heading is not (0.22–0.93), and the gap is dominated by omission rather than by wrong assertions. Released, uncampaigned variants. Four authored variants are released but not run, and no number in this paper derives from them: b_reroute and e_twisty (one and two closures with a detour of equal length), c_crossing (side-of-street tracking on), and d_vague (sign-reading off, geometry held; the only single-factor ablation in the benchmark). E.1 Step-cap censoring 215 of 450 Pathfinding conversations end at the 6,000-tick step cap and none at the wall-clock cap; per-system, per-channel shares are in Table 14. Pooled over cells, 88–100% of every system’s non-arrivals are step-cap endings rather than wrong destinations: across all 450 conversations only about a dozen non-arrivals are genuine wrong-destination errors, and GPT-Realtime-2.1-mini records none at all. The world is clock-limited, not accuracy-limited, and every zero in Table 14 is a lower bound rather than a measurement. On the realistic channel that carries the headline cells, 104 of 110 non-arrivals are step-cap endings and six are genuine wrong-destination errors. Nova 2 Sonic’s rate is a provider artefact: Amazon Bedrock terminates the session at its own cap regardless of task state, so its Pathfinding cells are censored by infrastructure as well as by behaviour. Appendix F Pathfinding in pictures One party holds the map and a telephone line. The other holds the street. Nothing else in Pathfinding is exotic, and the figures in this appendix are the fastest way to see what that premise does to a conversation. Figures 3–5 draw one scenario instance for each of the three Pathfinding conversation types, and each instance is drawn twice: once from a run that solved it and once from a run that did not. The two panels of each pair share the same start, the same destination and the same closures, so the contrast between them isolates the agent. Routes, closures and endpoints come from the stored simulation, and the outcomes stated in the captions are derived from the replay record rather than typed. The rerouting and all-day assistance instances are the same ones whose transcripts appear in Figures 2 and 6, so the excerpts there can be read against the routes here. Figure 3: single intent: one instance, two conversations. The walker starts at Madison Avenue and E 43rd Street facing east and wants Halstead Pharmacy at 3rd Avenue and E 43rd Street. Left: 3.1-Flash-Live walks the optimal five blocks and the call ends, arrived, in 147 seconds. Right: Realtime-2.1-mini, on the same instance, walks fifteen blocks, never arrives, and ends at the clock cap with its closest approach four blocks out. Figure 4: rerouting: one instance, two conversations. The walker starts at Madison Avenue and E 44th Street facing south and wants the Grand Central entrance on Lexington Avenue. Four pavement segments are shut, and none of them appear on the copilot’s map. Left: Voice Think Fast discovers the closures from the street, finds the surviving detour, and arrives after 19 walked blocks against an optimal seven given the closures. Right: Realtime-2.1, on the same instance, walks 18 blocks and ends at the clock cap without arriving. Figure 5: all-day assistance: one instance, two conversations. The walker starts at Madison Avenue and E 46th Street facing south. The first stop is Grand Central Market at Lexington Avenue and E 43rd Street, and only on arrival does the walker reveal the second, the United Nations visitor entrance at 1st Avenue and E 46th Street. Left: 3.1-Flash-Live completes both legs on the optimal eleven blocks. Right: Voice Think Fast reaches the first stop but never the second, and the call ends with the walker believing they had arrived, a failure the transcript alone would not reveal. Appendix G Acoustic channels, personas and distractors Both channels are telephony: G.711 μ-law at 8 kHz applies in the clean and the realistic preset alike, so clean means no additive degradation rather than an uncompressed signal. The realistic preset adds background noise mixed from a named recording inventory selected by task hash, burst noise events, frame drops, dynamic muffling and speech inserts, with LLM-driven interruptions and backchannels enabled; the clean preset disables both, so no decision model participates in clean runs. Because noise is mixed from recordings rather than synthesised at a target level, we report the file inventory with the release rather than a nominal SNR. Caller and walker speech is synthesised with ElevenLabs. Personas are pinned deterministically per scenario, so persona is a fixed property of a scenario and the same speaker serves it across all runs, gender-consistent with the authored caller. The control pool is American-accented and the regular pool is accent-diverse; the full persona inventory, with per-voice accent and speaking-rate attributes, ships with the release. The regular preset injects distractor events at 0.7 per minute (control: 0.0), drawn from vocal tics and non-directed phrases that are audible but carry no task content. At the corpus mean of 14.9 minutes per conversation this yields roughly ten injected events per call. Gold labels are written at injection time, which is what makes SELSEL scorable against ground truth and why it cannot be recovered from a control run that had nothing to ignore. user-simulator model metric luna sonnet-5 gem-3.6-fl TT 0.626±.075 0.716±.021 0.661±.048 CPCP 1.356±.167 1.644±.156 1.467±.189 SELSEL 0.439±.079 0.393±.079 0.378±.078 Pass@1Pass@1 0.378±.133 0.578±.144 0.267±.122 FAIFAI 2.044±.244 2.111±.222 1.733±.222 M^D M_D 3.343±.012 3.362±.011 3.338±.017 M^U M_U 4.077±.017 4.092±.013 4.066±.019 M^N M_N 3.564±.018 3.589±.016 3.550±.032 Table 10: User-simulator ablation. The agent is Realtime-2.1 in every arm; only the caller’s language model changes. n=45n=45 conversations per arm. The caller model moves Pass@1Pass@1 by 0.31, more than twice its widest interval here, and moves every dynamics metric; the MOS predictors barely move. Appendix H Metric definitions, judges and constants metric symbol range best turn-taking TT [0,1][0,1] high conversation progression CPCP 1–3 high selectivity SELSEL [0,1][0,1] high goal state GSGS 0,1\0,1\ high task reward, single run Pass@1Pass@1 [0,1][0,1] high task reward, at least 1 of 3 Pass@3Pass@3 [0,1][0,1] high task reward, all 3 of 3 Pass3Pass^3 [0,1][0,1] high over-effort ratio (share) ϱ+ ^+ (π+π^+) ≥1≥ 1 1 under-effort ratio ϱ− ^- ≤1≤ 1 1 under-effort share π−π^- [0,1][0,1] low faithfulness FAIFAI 1–3 high DNSMOS / UTMOS / NISQA M^D/M^U/M^N M_D/ M_U/ M_N 1–5 high Table 11: Metric glossary. The twelve-metric suite is the dynamics three (TT, CPCP, SELSEL), the agentic five (GSGS, Pass@1Pass@1, Pass3Pass^3, ϱ+ ^+, π−π^-) and the naturalness four (FAIFAI and the MOS predictors); Pass@3Pass@3 and ϱ− ^- are derived companions reported alongside. Definitions and formulas: Appendix H. Notation. A run yields a trajectory τ, a terminal world state sTs_T and the agent’s isolated audio channel y; [⋅]1[·] is the indicator; J∙J_ is an LLM judge on the stated scale. Conversational dynamics. For floor transfer t with offset δt _t (negative overlap, positive gap) and kind κt _t (clean handoff, agent barge-in, user barge-in, post-tool), with ϕκ:ℝ→[0,1] _κ:R→[0,1] the piecewise-linear on-time score and M(τ)M(τ) the count of user turns the agent never answered, T(τ)=[M(τ)=0]⋅1|T(τ)|∑tϕκt(δt),T(τ)=1\! [M(τ)=0 ]· 1|T(τ)| _t _ _t( _t), (1) ported verbatim from EVA-Bench (Bogavelli et al., 2026). CP(τ)=Jprog(τ)∈1,2,3CP(τ)=J_prog(τ)∈\1,2,3\ scores four binary dimensions (unnecessary tool calls, information loss, redundancy, question quality) with a topic-free prompt. SEL(τ)SEL(τ) is the fraction of injected distractors correctly ignored. Agentic capability. With h(⋅)h(·) a canonical hash of the record store and s⋆(x)s (x) the terminal state of a gold replay, GS(τ,x)=[h(sT)=h(s⋆(x))]GS(τ,x)=1[h(s_T)=h(s (x))]; in Pathfinding, GS=[vT=v⋆∧σT=σ⋆]GS=1[v_T=v _T=σ ] for terminal junction and side of street. ACT(τ,x)=[A⋆(x)⪯A(τ)]ACT(τ,x)=1[A (x) A(τ)] is the bipartite match of the gold action multiset against emitted tool calls on tool name and critical arguments (presence only: omission fails, surplus does not); NLA(τ,x)=Jassert(τ,x)NLA(τ,x)=J_assert(τ,x) judges the scenario’s natural-language assertion. The reward multiplies exactly the factors the scenario names, r(τ,x)=∏f∈B(x)f(τ,x)r(τ,x)= _f∈ B(x)f(τ,x) with B(x)⊆GS,ACT,NLAB(x) \GS,ACT,NLA\; in Pathfinding, r=GS⋅[η≥0.75]r=GS·1[η≥ 0.75] with η=ℓ⋆/ℓη= / the route-efficiency ratio. With cxc_x successes out of n=5n=5 runs of scenario x, Pass@1=1||∑xcxn,Pass3=1||∑x(cx3)/(53),Pass@1= 1|X| _x c_xn, ^3= 1|X| _x c_x3 / 53, (2) the latter τ-bench’s PasskPass^k at k=3k=3 (Yao et al., 2024). The at-least-once companion reported in Table 3 is Pass@3=1||∑x(1−(5−cx3)/(53))Pass@3= 1|X| _x (1- 5-c_x3 / 53 ), the probability that at least one of three draws passes. Effort is measured against the task’s reference workload: in the enterprise worlds ϱ(τ,x)=|A(τ)|/|A⋆(x)| (τ,x)=|A(τ)|/|A (x)| over the gold action list; in Pathfinding the ratio is over tool calls against a maneuver-and-lookup budget (ideal == optimal maneuvers + 3+\,3; replay.py), with the under-effort side reported as its share π−π^- alone. Route efficiency enters the reward only through η. Naturalness. FAI(τ)=mini≤5Jfaith(i)(τ)FAI(τ)= _i≤ 5J^(i)_faith(τ) over five binary dimensions judged with the agent’s instructions, role and tool schemas. DNSMOS OVRL, UTMOS22 and NISQA v2 mos_pred run on the isolated agent channel y. Pathfinding CPCP and FAIFAI are produced by the pipeline on a normalised [0,1][0,1] scale and mapped affinely (x↦1+2x 1+2x) onto 1–3 for reporting. judge metric model JprogJ_prog CPCP gpt-5.2 JfaithJ_faith FAIFAI claude-opus-4.6 JassertJ_assert NLANLA grok-4.3 Table 12: Judge models, served per metric via OpenRouter rather than as one stack. JassertJ_assert does not run in Pathfinding: every Pathfinding task has nl_assertions: null, consistent with the absence of an NLANLA factor in its reward. The judge prompts for JprogJ_prog, JfaithJ_faith and JassertJ_assert, and the piecewise-linear constants of ϕκ _κ, ship with the release. Appendix I Additional results system Pass@1Pass@1 ACTACT NLANLA no-tool WERcred (%) 2S_2S 0.013±0.008 0.041±0.015 0.104±0.030 0.527±0.038 33.3±4.2(n=242)^(n=242) 3.1_3.1 0.379±0.036 0.488±0.039 0.640±0.045 0.028±0.013 15.5±2.5(n=497)^(n=497) 2.1_2.1 0.480±0.039 0.655±0.039 0.741±0.044 0.061±0.019 10.4±2.2(n=488)^(n=488) mini_mini 0.207±0.030 0.370±0.043 0.437±0.053 0.052±0.016 20.2±2.7(n=490)^(n=490) VTF_VTF 0.481±0.038 0.589±0.037 0.682±0.045 0.041±0.015 14.4±2.0(n=495)^(n=495) pooled 0.312±0.015 0.427±0.018 0.520±0.022 0.142±0.012 17.1±1.2(n=2212)^(n=2212) Table 13: Enterprise deep dive, pooled across worlds and types: the single-run reward beside its non-state factors (ACTACT, NLANLA; 88 of 135 scenarios carry an assertion), the share of conversations with no tool call, and credential word error rate conditioned on episodes that attempted a credential (n in superscript). 385 episodes with no factor breakdown are excluded from ACTACT and NLANLA, which are therefore upper bounds; the exclusion is outcome-correlated and largest for mini_mini. No-tool is over all episodes; restricting to scenarios that require a tool moves 2S_2S to 0.463 and every other system by at most 0.006. clean realistic system Pass@1Pass@1 2 Sonic 0.00 0.00 0.00 0.00 0.00 0.00 3.1-Flash-Live 0.80 0.13 0.47 0.67 0.27 0.53 Realtime-2.1 0.53 0.07 0.33 0.33 0.00 0.27 Realtime-2.1-mini 0.33 0.07 0.00 0.13 0.00 0.13 Voice Think Fast 0.73 0.27 0.27 0.80 0.13 0.67 explore % 2 Sonic 50.0 47.8 50.0 66.7 53.3 41.2 3.1-Flash-Live 21.7 15.4 11.8 20.8 14.4 15.8 Realtime-2.1 41.1 38.5 30.6 50.0 47.2 34.8 Realtime-2.1-mini 38.8 36.8 25.1 47.3 34.5 34.0 Voice Think Fast 19.1 12.4 19.7 24.2 24.9 19.5 capped 2 Sonic 0.89 0.91 3.1-Flash-Live 0.18 0.00 Realtime-2.1 0.53 0.60 Realtime-2.1-mini 0.64 0.69 Voice Think Fast 0.22 0.11 Table 14: Pathfinding deep dive: Pass@1Pass@1 by type and channel (n=15n=15 per cell; a run passes if the walker arrived and η≥0.75η≥ 0.75), the exploration share (walker moves made under uncertainty, labelled per move from the tick stream), and the share of conversations ending at the step cap. 2 Sonic’s cap rate is dominated by its provider’s own session limit (Appendix E.1). Across all 450 conversations, 88–100% of every system’s non-arrivals are step-cap endings, so every 0.00 above is a lower bound. A score is a property of the world. GPT-Realtime-2.1 scores Pass@1Pass@1 of 0.200±.070 in Banking and 0.674±.081 in Travel: a range of 0.474, wider than the spread between the four engaged systems inside any single world (at most 0.445, in Healthcare). Yet the ordering is nearly frozen: Nova 2 Sonic is fifth and GPT-Realtime-2.1-mini fourth in all five enterprise worlds, while Grok Voice Think Fast 1.0, GPT-Realtime-2.1 and Gemini-3.1-Flash-Live fill the top three, Grok Voice Think Fast 1.0 leading in three worlds and GPT-Realtime-2.1 in two, by margins their intervals mostly do not resolve. A reported completion score is therefore about as much a statement about the world it was measured in as about the system, and the interval a system spans across the five enterprise worlds is the reference we read everything else against. Pathfinding sharpens the point: on the four metrics computed by identical code everywhere, 15 of 20 system–metric cells fall outside that reference interval (Appendix J), while on the reward no system leaves its own five-world range. Judged on completion alone, Pathfinding behaves like a sixth subject; what breaks is how the conversation goes. Reliability is a property of the scenario. Read along the solid curves of Figure 1, Grok Voice Think Fast 1.0 retains 54% of its Pass@1Pass@1 at k=3k=3, GPT-Realtime-2.1 49%, Gemini-3.1-Flash-Live 41% and GPT-Realtime-2.1-mini 30%, so a single-run leaderboard overstates what any system will do three times running; by k=5k=5 the best system holds 0.154, while under retries the top three reach 0.72–0.77. Successes are not i.i.d.: every engaged enterprise cell exceeds the Pass@13Pass@1^3 prediction of Pass3Pass^3, by 1.2×1.2× ( GPT-Realtime-2.1, Travel) up to 91×91× ( GPT-Realtime-2.1-mini, Healthcare). Failure concentrates on particular scenarios rather than spreading across draws, so a system’s failures are diagnosable rather than stochastic; the strongest-to-weakest engaged ratio widens from 2.6×2.6× at k=1k=1 to 4.8×4.8× at k=3k=3. No single column summarises a system. Nova 2 Sonic posts the best selectivity in every world (0.960–0.995) while reaching the full reward in at most 3.7% of conversations and returning the lowest conversation progression everywhere (1.007–1.030, against 1.156–2.207 for the other four); GPT-Realtime-2.1-mini sits at or near the top of the UTMOS and NISQA columns while placing fourth of five on the reward everywhere. The two columns that come closest to a summary, Pass@1Pass@1 and Pass3Pass^3, track each other at ρ=+0.96ρ=+0.96 and everything else much less. The clearest single collapse. Nova 2 Sonic’s turn-taking falls from 0.611–0.760 in the enterprise worlds to 0.062±.058 in Pathfinding, while its selectivity is unchanged (0.991), it speaks in only 12.3% of the call, its first tool call comes at 48.4 s against 15–27 s for the other four, and it under-works the task in 73% of conversations. The system did not change its behaviour; the world changed what that behaviour costs. In a support call an agent can decline to act and still answer, because the caller waits. In Pathfinding the walker keeps moving and keeps reporting, unanswered turns accumulate, and the unanswered-turn term zeroes the score. Pathfinding is the only world in DuplexWorld where turn-taking is not winnable by silence, and the only one where the disengaged system is last on every pillar. One caveat accompanies the cell: Nova 2 Sonic’s Pathfinding sessions are additionally censored by its provider’s own session cap (Appendix E.1). Reaching the right state is not doing the task. Across the 25 enterprise cells the mean of GS−Pass@1GS-Pass@1 is +0.255+0.255, and every cell is positive: largest in Banking (+0.381+0.381 on average, up to +0.489+0.489) and smallest in Logistics (+0.093+0.093). The remaining conjuncts, acting through the sanctioned sequence and saying what was done, are where the reward is lost (Tables 13 and 18 decompose them). A do-nothing agent passes GSGS wherever the gold terminal state equals the seeded state, and Nova 2 Sonic makes this visible: gold state in 34.1–41.5% of conversations in four worlds while issuing no adequate tool call in most, against 5.2% in Logistics, the world with the least to refuse. GSGS is a capability measure only conditioned on the world containing something to change, and a benchmark that headlines goal state is partly reporting how many of its tasks are no-ops. Why accuracy and experience trade off. The systems that call tools most have the worst turn-taking: Gemini-3.1-Flash-Live falls short of the reference workload in 10% and 3% of Logistics and Travel conversations and scores 0.443 and 0.314 on TT, the worst in both, while Nova 2 Sonic falls short in 77% and 64% and scores 0.631 and 0.695. Tool calls introduce silence, silence is scored as a timing failure, and any composite that gates on timing penalises acting. This gives a mechanism for EVA-Bench’s finding that no system exceeds 0.5 on both accuracy and experience (Bogavelli et al., 2026). How much of the score is the harness? Three configuration dials, none of which any published voice benchmark reports, move scores by as much as the choice of system does (Appendix L). At the default 0.5 server-VAD threshold, some synthesised personas are effectively inaudible: one triggered the agent in 0 of 27 sessions, another passed 1 of 77 runs, and the agent greeted and then fell silent in 67 of 128 simulations; at 0.2 the effect largely disappears. A benchmark pairing synthetic personas with a default VAD threshold is partly measuring persona loudness, and the resulting silence is indistinguishable from an agent that chose not to speak. Holding the agent fixed, the choice of user-simulator model moves Pass@1Pass@1 from 0.267 to 0.578 (Table 10), while the MOS predictors barely move, consistent with their measuring the agent’s audio rather than the interaction. And the step cap force-zeroes runs that do not terminate: excluding capped runs moves one system’s pooled score from 0.383 to 0.469 and another’s from 0.108 to 0.213, narrowing a reported 3.5×3.5× system gap to 2.2×2.2×. Inside Pathfinding: the mechanisms bite. rerouting costs every system most of its reward. No system exceeds Pass@1Pass@1 of 0.27 in either channel, against up to 0.80 on single intent, while the acoustic channel moves Pass@1Pass@1 far less than the geometry does, and in no consistent direction ( Gemini-3.1-Flash-Live rises from 0.13 to 0.27 under realistic acoustics on rerouting; Grok Voice Think Fast 1.0 falls from 0.27 to 0.13). all-day assistance sits between, and only the systems that survive rerouting recover on it. The reward’s operating point matters less than it might: moving the efficiency threshold η0 _0 across [0.5,1.0][0.5,1.0] moves levels far more than ranks, the one rank event being a 0.044 swap between the top two at the operating point itself (Table 9, Appendix E). Exploration separates systems, not conditions. The exploration share is this world’s between-system marker: 16.2% of Gemini-3.1-Flash-Live’s walker moves (move-weighted) and 22.7% of Grok Voice Think Fast 1.0’s are made under uncertainty rather than on a confident instruction, against 40.3% for GPT-Realtime-2.1-mini, 43.4% for GPT-Realtime-2.1 and 51.2% for Nova 2 Sonic, and across the five systems it tracks the reward at ρ=−0.80ρ=-0.80. Within a system, moving to a harder condition does not raise it consistently, so it marks systems rather than conditions, and we make no causal claim. It gives the efficiency story its mechanism: GPT-Realtime-2.1 is efficient and still lost (ϱ+ ^+ of 1.48 over 40% of conversations, below its own five-world range, with GSGS 0.400 also below), while Nova 2 Sonic wanders (ϱ+ ^+ of 4.25, on a conditioning population of eight conversations). Efficient-and-lost and inefficient-and-lost are different failures, and the exploration share separates them where the reward alone cannot. Almost none of these failures is a wrong destination: on the realistic channel, 104 of 110 non-arrivals end at the step cap and only six end at the wrong place (Appendix E.1). Nobody gets lost; they run out of time. Appendix J Range analysis Table 15 reads every Pathfinding value against the interval the same system spans across the five enterprise worlds: 15 of the 20 cells on the four identical-code metrics leave that interval, no Pass@1Pass@1 cell does, and the goal-state analogue leaves it for all five systems. identical code analogue system TT CPCP SELSEL FAIFAI GSGS Pass@1Pass@1 2 Sonic 0.062⇓ 1.022 0.991 1.244⇓ 0.000⇓ 0.000 3.1-Flash-Live 0.538⇑ 1.178⇓ 0.782⇑ 1.511 0.978⇑ 0.489 Realtime-2.1 0.561 1.289⇓ 0.551⇑ 1.644⇓ 0.400⇓ 0.200 Realtime-2.1-mini 0.387⇓ 1.156⇓ 0.526⇑ 1.333⇓ 0.311⇓ 0.089 Voice Think Fast 0.682 1.378⇓ 0.507⇓ 1.489⇓ 0.867⇑ 0.533 outside range 3/5 4/5 4/5 4/5 5/5 0/5 Table 15: Pathfinding against each system’s five-world range. ⇑ /⇓ mark values above/below the interval the system spans across the five enterprise worlds (Table 3); unmarked values fall inside it. The left block is computed by identical code in every world; the right block substitutes the world’s terminal state, so part of any difference is definitional. The move to navigation lands on how the conversation goes, not on whether the task is completed. Appendix K Metric–reward associations Table 16 gives the Spearman rank correlation of every metric with Pass@1Pass@1 over the 25 enterprise world×system cells, and over the 20 cells that exclude the disengaged system. Turn-taking is uncorrelated with the reward (ρ=−0.06ρ=-0.06) and selectivity negatively correlated (−0.45-0.45, collapsing to +0.05+0.05 once the disengaged system is excluded); conversation progression (+0.83+0.83), the goal-state factor (+0.77+0.77) and faithfulness (+0.58+0.58) track it, and none of these can be raised by abstaining. The correlations do not show the abstention metrics are useless among engaged systems. They show one disengaged system is enough to top a leaderboard built on them, and a metric a system can win by not working is broken whether or not anyone exploits it. The exception is the under-effort share π−π^-, the strongest process predictor of success in the suite (ρ=−0.85ρ=-0.85; −0.73-0.73 engaged-only): a measure of inactivity that does not reward inactivity, because it is signed against the task’s own reference workload. The MOS predictors go flat to negative among the engaged four, consistent with Section 5.1. metric all 25 cells engaged 20 Pass3Pass^3 +0.96+0.96 +0.93+0.93 CPCP +0.83+0.83 +0.69+0.69 GSGS +0.77+0.77 +0.61+0.61 FAIFAI +0.58+0.58 +0.63+0.63 M^U M_U +0.49+0.49 +0.02+0.02 M^N M_N +0.29+0.29 −0.37-0.37 M^D M_D −0.04-0.04 −0.39-0.39 TT −0.06-0.06 +0.26+0.26 SELSEL −0.45-0.45 +0.05+0.05 π−π^- −0.85-0.85 −0.73-0.73 Table 16: Metric–reward associations: Spearman rank correlation of each metric with Pass@1Pass@1 over the 25 enterprise world×system cells (Table 4 for π−π^-), and over the 20 cells excluding the disengaged system. These are associations between our own metrics, not against human ratings. Appendix L Harness dials Table 17 collects the three configuration dials beside the system-identity contrast; the closing paragraph of Appendix I reads them. dial levels Δ Pass@1Pass@1 VAD threshold 0.2 / 0.5 personas to full silence step cap scored / excluded +0.086+0.086, +0.105+0.105 simulator LM 3 models 0.267 – 0.578 system identity 5 systems 0.00 – 0.67 Table 17: Harness dials and effect sizes, with system identity as the reference contrast; the simulator-LM row is Table 10; the VAD row is the persona-audibility effect of Section I (personas inaudible at the 0.5 default); the step-cap row is the shift in two systems’ pooled Pass@1Pass@1 when capped runs are excluded rather than scored zero. Appendix M Duplex phenomenon audit We print this audit rather than leave the claim on the title page unsupported: a benchmark’s honest description includes the gap between what it elicits and what it scores. The harness elicits the full-duplex phenomena (user-side barge-in at a fixed cadence, backchannels, overlapping speech, dead air, and mid-utterance revision) and the suite scores a subset: TT scores every floor transfer’s offset and zeroes on unanswered turns, SELSEL scores the response to every injected distractor, and the suspension type scores dead air across the gap. Elicited but not scored: backchannel quality and placement, the prosody of overlapping speech, and whether an agent’s mid-utterance revision is fluent or abrupt. The reward never sees any of these, so a system could be graceless at every overlap and lose nothing on the agentic pillar, which is why dynamics stays a pillar beside the reward rather than folded into it, and why no composite is formed. Appendix N Conversation-type excerpts: failure modes Figure 6: Characteristic failure modes: the failing analogues of Figure 2, drawn from failing runs of the scored corpus, with the mechanism of each failure stated beneath the excerpt. The two Pathfinding tiles are the same task instances as their successes in Figure 2. Appendix O Enterprise results by conversation type metric system Pass@1Pass@1 2S_2S 0.000 0.000 0.000 0.000 0.120 0.000 0.000 0.000 0.000 3.1_3.1 0.493 0.320 0.360 0.347 0.307 0.440 0.467 0.320 0.360 2.1_2.1 0.640 0.533 0.573 0.440 0.507 0.413 0.440 0.333 0.440 mini_mini 0.320 0.107 0.160 0.267 0.227 0.187 0.200 0.107 0.293 VTF_VTF 0.680 0.493 0.453 0.360 0.413 0.467 0.667 0.440 0.360 ACTACT 2S_2S 0.000 0.000 0.100 0.000 0.155 0.000 0.042 0.000 0.069 3.1_3.1 0.575 0.342 0.568 0.478 0.333 0.559 0.542 0.557 0.437 2.1_2.1 0.841 0.597 0.671 0.750 0.582 0.625 0.567 0.629 0.651 mini_mini 0.475 0.192 0.492 0.435 0.297 0.294 0.386 0.250 0.492 VTF_VTF 0.720 0.569 0.613 0.561 0.389 0.662 0.667 0.606 0.500 NLANLA 2S_2S 0.000 0.000 0.029 0.191 0.288 0.174 0.000 0.095 0.000 3.1_3.1 0.875 0.400 0.533 0.603 0.739 0.833 0.850 0.492 0.635 2.1_2.1 0.811 0.684 0.685 0.774 0.828 0.857 0.667 0.623 0.762 mini_mini 0.515 0.444 0.210 0.596 0.542 0.409 0.375 0.391 0.450 VTF_VTF 0.825 0.778 0.627 0.612 0.712 0.750 0.850 0.556 0.731 no-tool 2S_2S 0.560 0.293 0.613 0.547 0.693 0.493 0.520 0.467 0.560 3.1_3.1 0.013 0.000 0.013 0.000 0.080 0.093 0.000 0.027 0.027 2.1_2.1 0.000 0.013 0.147 0.000 0.200 0.133 0.013 0.013 0.027 mini_mini 0.013 0.013 0.107 0.040 0.080 0.187 0.000 0.000 0.027 VTF_VTF 0.000 0.000 0.040 0.000 0.160 0.133 0.000 0.013 0.027 WERcred (%) 2S_2S 39.7(25) 25.0(40) 36.4(22) 41.5(29) 40.4(15) 32.8(29) 30.1(28) 29.0(31) 33.7(23) 3.1_3.1 14.4(59) 10.5(55) 17.7(49) 18.0(65) 25.2(52) 12.0(49) 12.6(55) 11.7(54) 16.9(59) 2.1_2.1 7.8(60) 3.0(55) 3.5(43) 17.3(65) 18.7(49) 23.0(49) 7.1(54) 2.3(54) 10.5(59) mini_mini 11.9(60) 19.4(54) 23.2(47) 19.8(62) 32.7(49) 27.5(50) 22.9(55) 14.7(55) 13.1(58) VTF_VTF 14.2(60) 11.6(55) 18.4(49) 16.2(65) 18.0(49) 9.2(49) 11.3(55) 11.0(54) 18.9(59) Table 18: Enterprise deep dive by conversation type, pooled over the five enterprise worlds (episode-weighted; WERcred conditioned on credential-attempting episodes with cell n in superscript). ACTACT and NLANLA inherit the outcome-correlated exclusion of Table 13 and are upper bounds. 2S_2S’s WERcred row is conditioned on the minority of its calls in which it acted at all. Table 18 decomposes the enterprise deep dive of Table 13 by conversation type. Appendix P Run statistics The scored corpus comprises 3,825 conversations (3,375 enterprise, 450 Pathfinding), totalling 387 hours of simulated speech. The Pathfinding block is 9 campaigned scenarios × 5 systems × 5 runs × 2 channels, median duration 20.0 minutes; 215 of its 450 conversations end at the step cap and none at the wall-clock cap (Appendix E.1), and its realistic channel carries the headline cells (225 conversations). In the enterprise worlds, 302 conversations end at the wall-clock timeout; every capped or timed-out run scores reward 0 and remains in the corpus (Section 4.2). Appendix Q Results in pictures Figures 7–10 redraw the headline numbers of Section 5 for visual comparison. Figure 7: Pass@1Pass@1 by world and pooled overall; whiskers are 95% bootstrap intervals. Figure 8: Share of conversations issuing more (solid, up) or fewer (faded, down) tool calls than the reference workload. Figure 9: All eleven metrics of Table 3, pooled over worlds and min–max scaled per axis; larger is better (SEL inverted). Sectors mark the three pillars. Figure 10: Exploration share in Pathfinding by system and conversation type (realistic channel); whiskers are 95% bootstrap intervals. Appendix R License of Artifacts All artifacts are used for non-commercial academic research, consistent with their terms. The five evaluated systems are proprietary services accessed through their providers’ APIs under the corresponding terms of service: Nova 2 Sonic via Amazon Bedrock, Gemini-3.1-Flash-Live via the Google Gemini API, GPT-Realtime-2.1 and GPT-Realtime-2.1-mini via the OpenAI API, and Grok Voice Think Fast 1.0 via the xAI API. The harness-side models are likewise proprietary services: the user simulator gpt-5.6-luna via Azure, and the decision model claude-haiku-4.5 and the judge models (gpt-5.2, claude-opus-4.6, grok-4.3) via OpenRouter under each provider’s terms; caller and walker speech is synthesised through the ElevenLabs API under its terms. No provider model weights are redistributed. Among open-source evaluation components, the τ-bench and τ2τ^2-bench harnesses our orchestrator descends from are MIT-licensed, as is the DNSMOS code in the DNS-Challenge repository and the UTMOS (UTMOS22) system; NISQA is MIT-licensed code with pretrained weights under C BY-NC-SA 4.0, which our non-commercial use respects; the turn-taking metric is reimplemented from EVA-Bench’s published specification (Bogavelli et al., 2026). Telephony coding follows the public ITU-T G.711 standard. All scenarios, records, policies, personas and identifiers in DuplexWorld are authored by us and contain no third-party data, and the realistic channel’s background-noise recording inventory is distributed with the release. We release the scenario corpus, annotations and evaluation code for research use, under C BY 4.0 for data and MIT for code. Appendix S Use of AI Assistants AI assistants were used in the writing and coding process, and not as part of the research itself. In writing, assistance was limited to grammar and spell checking. In coding, an AI-assisted IDE (Claude Code) was used during implementation. Task design, scenario authoring, experiments, analyses and conclusions are the authors’ own.