> ## Content Index > Fetch the complete content index at: https://www.notatechguy.com/llms.txt > Use this file to discover other available public pages before exploring further. # Voice agents drop 9.7% when instructions are merely implied - URL: https://www.notatechguy.com/voice-agents-drop-9-7-when-instructions-are-merely-implied/ - Published: 2026-09-04T15:24:08.000Z - Updated: 2026-09-04T15:24:08.000Z - Description: DSB-IFEval tests 1,038 cases across eight assistant roles and finds voice agents drop up to 9.7% when instructions are implied, not stated. - Author: Marcello Babbili - Tags: Technology & AI, AI Models, OpenAI A new benchmark with 1,038 test cases across eight assistant roles finds that full-duplex voice agents, the kind that listen and speak at the same time, lose up to 9.7% adherence when conversational rules are implied by their persona rather than stated outright [S¹](https://arxiv.org/abs/2609.03423?ref=notatechguy.com). The gap exposes a problem the voice AI industry has been able to ignore: models that sound like a character but cannot act like one. What breaks first, and who pays for it, depends on a distinction most benchmarks have never bothered to test. **My read:** This is the first benchmark I've seen that separates what a voice agent says from how it behaves in real time. Most voice AI evaluation checks whether the model produces the right words. DSB-IFEval checks whether it knows when to interrupt and when to break character for safety. The 9.7% drop for F-Actor under persona-only conditioning is not a cross-model ranking; it is a within-model sensitivity signal, and a 9.7% swing in floor management is enough to make a customer-service agent feel rude or timid in live conversation. I don't buy the idea that any tested system is good at implicit instruction following yet. The paper itself says the best content-adherent models, including GPT-Realtime, have floor behavior that does not adapt. That is the finding that matters. ## What the benchmark actually tests DSB-IFEval, published on arXiv on 4 September 2026, evaluates implicit instruction-following in real-time spoken interaction [S¹](https://arxiv.org/abs/2609.03423?ref=notatechguy.com). The name is dense but the idea is simple: when you tell a voice agent "you are a stern tutor," does it figure out on its own that it should interrupt rambling students, or does it need an explicit rule saying "interrupt when the student goes off-topic"? The benchmark runs 1,038 test cases across eight assistant roles [S¹](https://arxiv.org/abs/2609.03423?ref=notatechguy.com). Each case applies one of five conditioning protocols [S¹](https://arxiv.org/abs/2609.03423?ref=notatechguy.com): - Default behavior: no special instructions, just the base system. - Explicit behavioral instructions: the rule is stated directly. - Persona-implied behavior: the rule is never stated; the model must infer it from the role description. - Combined persona-rule conditioning: both the persona and the explicit rule are given. - Instruction conflict: two instructions contradict each other, and the model must resolve which one wins. Two scores measure performance. The Instruction Adherence Score (IAS) is deterministic and tracks real-time floor management, meaning whether the agent interrupts or yields at the right moments [S¹](https://arxiv.org/abs/2609.03423?ref=notatechguy.com). The Persona Adherence Score (PAS) uses an LLM as judge to score whether the content the agent produces is consistent with its assigned character [S¹](https://arxiv.org/abs/2609.03423?ref=notatechguy.com). Prior work has tested instruction-following in text models [P⁵](https://arxiv.org/abs/2311.07911v1?ref=notatechguy.com), in speech-aware language models [P³](https://arxiv.org/abs/2505.19037v1?ref=notatechguy.com), and in multi-turn duplex dialogue [P⁴](https://arxiv.org/html/2510.07838v1?ref=notatechguy.com). A separate line of work, Instruct-FD, examined turn-taking instructions specifically [P²](https://arxiv.org/html/2607.20460?ref=notatechguy.com). DSB-IFEval is the first to test whether models can infer behavioral rules from persona alone, rather than being told. ## The gap between sounding right and acting right The headline finding is the gap between explicit and implicit conditioning. F-Actor, a full-duplex model, drops 9.7% in instruction adherence when rules are implied by persona alone rather than stated explicitly [S¹](https://arxiv.org/abs/2609.03423?ref=notatechguy.com). PersonaPlex, another full-duplex system, drops 4.5% under the same conditions [S¹](https://arxiv.org/abs/2609.03423?ref=notatechguy.com). ![Adherence drop under persona-only conditioning (within-model)](https://storage.ghost.io/c/6e/89/6e896869-22ef-4281-a213-b4c462c17cff/content/images/2026/09/chart_a13a38c59ce1df3cfd2d.png) These are within-model comparisons, not a head-to-head ranking. They measure how much each model degrades when it has to read between the lines instead of following a direct order. A separate cluster of models, GPT-Realtime, MiniCPM-o, and Fun-Audio-Chat, tells a different story. These systems produce persona-consistent content reliably, but their floor behavior does not change across explicit and persona-only instructions [S¹](https://arxiv.org/abs/2609.03423?ref=notatechguy.com). They say the right things. They do not adjust when they say them. Several proactive actions, like voluntary interruption, remain constrained. DSB-IFEval adds a layer: even when the voice channel works, the model may not grasp the social rules a human would pick up from context. ## The safety override problem The most unsettling finding sits in the conflict protocol. Systems that reliably follow conflicting directives tied to their prescribed persona still struggle to override those directives under safety conflict [S¹](https://arxiv.org/abs/2609.03423?ref=notatechguy.com). In plain terms: if a persona implies the agent should always push back, and a user says something that triggers a safety rule, the agent may not know which instruction wins. This is not a hypothetical. A customer service agent told to be persistent might refuse to drop a harmful line of conversation. A medical assistant told to be reassuring might soften a safety warning. The benchmark shows the models can follow persona directives but cannot reliably break them when safety demands it. ## What to do about it For teams building voice agents, the takeaway is to stop relying on persona descriptions to carry behavioral rules. If you need an agent to interrupt after three seconds of silence, say so explicitly. The benchmark shows a measurable drop when rules are only implied. Consider a triage nurse line that deploys a voice agent to screen incoming calls. The persona description says "calm and efficient." That implies the agent should redirect rambling callers back to their symptoms. But under persona-only conditioning, the model may not interrupt at all. The fix is an explicit rule: "If the caller has not described a symptom within 15 seconds, ask them to state their primary concern." The benchmark data suggests this closes most of the adherence gap. One practical step this week: audit your voice agent prompts for any behavioral rule that exists only as a persona implication. List each one. Rewrite it as an explicit instruction. The 9.7% drop in the benchmark is the ceiling of what you lose by leaving rules implicit. ## What we don't know yet The paper is an arXiv preprint and has not been peer-reviewed [S¹](https://arxiv.org/abs/2609.03423?ref=notatechguy.com). The PAS score relies on an LLM as judge, which can introduce evaluator bias or inconsistency. The tested systems represent specific versions and APIs; current capabilities may differ from what was benchmarked. The benchmark does not cover all commercially available full-duplex voice agents, and the paper does not identify a single best architecture. The 9.7% and 4.5% figures describe within-model sensitivity, not a performance ranking between F-Actor and PersonaPlex. The next signal: any API update to GPT-Realtime or the other tested systems. We will recheck the DSB-IFEval scores against new model versions to see whether the explicit-to-implicit gap closes. If you want reporting like this in your feed every week, subscribe. We follow the benchmarks that expose what AI actually does, not what the demo videos promise. --- *Sources: [S1 — DuplexSpeechBench-IFEval: Evaluating Implicit Instruction Following in](https://arxiv.org/abs/2609.03423?ref=notatechguy.com) · [P2 — Instruct-FD: Can Your Full-Duplex Speech System Follow Turn-Taking Ins](https://arxiv.org/html/2607.20460?ref=notatechguy.com) · [P3 — \[2505.19037v1\] Speech-IFEval: Evaluating Instruction-Following and Qua](https://arxiv.org/abs/2505.19037v1?ref=notatechguy.com) · [P4 — Full-Duplex-Bench-v2: A Multi-Turn Evaluation Framework for Duplex Dia](https://arxiv.org/html/2510.07838v1?ref=notatechguy.com) · [P5 — Instruction-Following Evaluation for Large Language Models](https://arxiv.org/abs/2311.07911v1?ref=notatechguy.com)* ## More from Not A Tech Guy - [When to trust LLM recommendations: four-tier framework](https://www.notatechguy.com/when-to-trust-llm-recommendations-four-tier-framework/) - [New technique cuts AI agent wait time up to 45%](https://www.notatechguy.com/new-technique-cuts-ai-agent-wait-time-up-to-45/) - [GPT-6 Astra crosses OpenAI's Critical cybersecurity threshold](https://www.notatechguy.com/gpt-6-astra-crosses-openai-s-critical-cybersecurity-threshold/) --- *Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.*