acceptodds
Withdrawn from ICLR 2027

Streaming Intent Transactions for Safe Asynchronous Tool Use in Full-Duplex Voice Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Full-duplex voice agents can dispatch tools before a user finishes speaking, reducing latency but creating a temporal authorization problem: a callback may return after the user has revised or revoked the intent that authorized it. Existing evaluations primarily measure routing and tool correctness rather than whether a delayed result remains authorized to change external state. We formalize this failure mode as streaming intent transactions and propose DuplexTxn. A causal policy over a frozen Omni backbone compares dispatch- and commit-time representations, while a deterministic gate verifies version, binding, dependency, confirmation, and idempotency constraints. We also introduce DuplexTxn-Bench, a paired-timing evaluation that holds speech, tool output, and initial state fixed while moving callback arrival across a spoken revision boundary. Across 516 synthetic out-of-distribution families and four independently trained seeds, DuplexTxn reduces unsafe-effect episodes from 17.05% to 0.024%, eliminates observed stale-result effects, and raises safe task completion from 8.58% to 82.63% relative to our execution of released DuplexOmni. On a post-freeze natural-speech transfer panel, the same frozen policies reach 100% safe completion under controlled callback replay. These results support temporal authorization as a distinct requirement for safe full-duplex tool use.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
full-duplex voice agents, asynchronous tool use, streaming intent tracking, transaction safety, multimodal agents

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_fnmKOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies the temporal authorization problem in full-duplex voice agents: users may change their minds after authorizing a tool call. The authors argue that a tool call should be delayed or dropped if the user's intent changes. They propose DuplexTxn, a policy that compares dispatch-time and commit-time representations and uses a gate to determine how to handle issued tool calls. They also introduce DuplexTxn-Bench to evaluate such scenarios. The authors implement the DuplexTxn policy on DuplexOmni, and the results show significant improvements in model performance on the target scenarios.

Critical strengths
  • The problem studied in this paper could arise in real-world applications of full-duplex voice agents.
  • The proposed method demonstrates significant performance improvements over the baseline.
Critical weaknesses
  • Questionable Setting: I think the biggest problem with this paper lies in its setting. The paper targets scenarios in which a user authorizes a tool call but later regrets the decision and attempts to revoke it. The core assumption of this setting is that there is enough time to revoke the tool call and that the tool call itself is revocable. However, sometimes tool calls are executed so quickly that there is insufficient time to revoke them. In other cases, tool calls are irreversible. For example, once a user authorizes a payment, the transaction may be processed instantly and cannot simply be revoked. Instead, it must be refunded, which requires a completely different tool call and may take considerable time. Therefore, I personally think that such scenarios should be handled differently: if a tool call has already been executed and the user wants to cancel it, the agent could issue another tool call to reverse its effects, if possible. For sensitive operations (such as flight bookings), the agent could explicitly ask the user for confirmation a second time and wait for final authorization.

  • Unclear Tool Call Decision Time: The core of the proposed method is the commit/wait/drop decision. However, there is no clear explanation of how long a tool call should remain in the waiting state before being committed or dropped. This is not merely about reporting a specific number; it concerns the fundamental design philosophy of the method. For example, if a tool call is delayed for too long, latency increases significantly. If the waiting period is too short, the user may change their mind after the tool call has already been committed. Therefore, it is unclear how the proposed method addresses this fundamental trade-off.

  • Critical Details Missing:

    • Transaction States: The authors design 41 transaction states, but do not provide a comprehensive list of them.
    • Benchmark Construction: Important details about the benchmark are missing. For example, what tools are included in the benchmark? How are the benchmark cases generated? There is no clear description of the construction pipeline or algorithm. Do the authors use LLMs or TTS systems to construct the benchmark?
    • Unclear Metric Definitions: The paper does not provide rigorous mathematical definitions for all evaluation metrics, leaving several ambiguities. For example, what exactly counts as a "task predicate" in the Safe Completion metric? For the Missed Effect metric, over what time window is it calculated?
Questions for authors

N/A

Reviewer_Pwk3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies temporal authorization in full-duplex voice agents: whether a delayed tool result remains authorized to change external state after the user revises or revokes the authorizing intent. The proposed method, DUPLEXTXN, combines causal intent tracking, comparison of dispatch-time and commit-time representations, and an atomic commit gate. DUPLEXTXN-Bench evaluates this capability through a paired-timing design that varies callback arrival while holding speech, tool output, and initial state fixed.

Critical strengths
  • The paper separates semantic intent judgments from ledger-relative runtime constraints and clearly explains the scope of its deterministic guarantees and the remaining risks from learned-model errors.
  • The paired-timing design isolates the effects of callback timing while measuring both safety and task completion. Table 8 also reports an approximately 10.3 percentage-point improvement in safe task completion over local Cordon- and Atomix-inspired transaction controls receiving the same causal semantic events.
Critical weaknesses
  • The natural-speech results do not establish a favorable tradeoff among safety, task completion, and latency. In the native-call diagnostic, DUPLEXTXN achieves 19.30% safe task completion, compared with 29.53% for the turn-final latest-call-wins control, while reducing stale-effect episodes from 9.40% to 6.54% (Section 6.3). Without a direct end-to-end latency comparison on this panel, it remains unclear whether the benefit of earlier decisions justifies the lower task completion rate.
Questions for authors
  • How much end-to-end latency does DUPLEXTXN save relative to the turn-final latest-call-wins control on the native-call panel, and does this benefit justify the observed reduction in safe task completion?
Reviewer_VsXgOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies temporal authorization in full-duplex voice agents: a tool call may be authorized when dispatched but become stale before its result returns because the user has corrected or revoked the request. It proposes DUPLEXTXN, which combines a lightweight causal policy over a frozen Omni backbone, a dispatch–commit representation comparison that selects COMMIT/WAIT/DROP, and a deterministic gate enforcing version, binding, dependency, confirmation, and idempotency constraints.

The main evaluation contribution is a paired-timing benchmark that holds speech, tool results, and initial state fixed while moving callback arrival across a spoken revision boundary. The method greatly increases safe completion and greately reduces unsafe episodes relative to the authors’ execution of released DuplexOmni.

Critical strengths

I like the setting of the question. The benchmark isolates a consequential failure mode and evaluates safety alongside utility. The separation between learned semantics and deterministic enforcement is useful, and the ablations support a contribution beyond simply blocking writes.

Critical weaknesses

Figure 1 requires the system to hold a hotel callback during an unresolved repair, then preserve the hotel transaction while revoking only the flight transaction. However, the author explicitly states that parallel requests, backchannels, and unrelated speech have no positive training support. The author describes a preserve/invalidate head over the active mandate, allows multiple calls to share a version, and states that invalidation revokes older lineage. It does not explain how the model identifies which of several live transactions should be invalidated, or how the single active-intent notation relates to the separate hotel and flight version.

Clarification of task scoring: What are the exact final-state completion predicates for the two timing arms of the controlled panel? After a revision, must the revised-value effect execute, or is dropping the pre-initialized old-value callback sufficient? How is an authorized pre-revision commit scored once the subsequent revision occurs?

An agent that delays commitment can observe additional speech and avoid actions that a prompt agent would legitimately execute and subsequently need to compensate for. Thus, differences in final-state completion may reflect commitment timing and exposure to recovery costs, in addition to authorization reasoning. Holding callback schedules fixed does not by itself establish matched commitment times when WAIT is available, and the equal P95 latency reported for one ablation does not resolve this issue for the main comparison.

There is no place to find the dataset queries or the training codebase.

Questions for authors

See above.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.