Model confirms details back instead of calling the tool, even when told not to

#3
by vahidkowsari1 - opened

Title

Model confirms details back instead of calling the tool, even when told not to

Body

Sharing a reproducible behavior we hit while evaluating alpha-1 for a phone agent. On flows where a tool call must happen before the conversation can proceed, PhoneLLM reliably spends the turn reading the caller's details back and asking for confirmation, instead of calling the tool.

Minimal reproduction

Three-line system prompt, exactly one tool, an explicit instruction to call it immediately, temperature=0, thinking disabled.

System prompt

You are a phone assistant for a medical clinic.
To look up a patient record you must first verify identity.
When the caller has given first name, last name and date of birth, call the
verify_identity tool immediately. Do not ask for anything else.

Tool — a single verify_identity function taking first_name, last_name, date_of_birth.

User turn

Hi, I'd like to look at my account. My name is Jordan Avery, date of birth March 20, 1985.

All three required arguments are present in that one utterance.

PhoneLLM Alpha 1 — no tool call, 3/3 samples:

"I can certainly help you with that. Just to confirm, that's Jordan Avery,
 date of birth March 20, 1985. Is that correct?"

gpt-4.1-mini, same prompt, same tool, same user turn — tool call, 3/3 samples:

tool_calls: [verify_identity]
content: ""

The behavior is unchanged at temperature=0 and 0.1, with 1 tool or 27, and with a short prompt or a large production one. Adding the anti-narration rules from the pipecat-examples/phonellm bot ("don't narrate your own tool use", "Either answer, or call the tool and then say what it did") removed the trailing question mark and nothing else.

Why it matters for gated flows

The example bot is a reservation flow, where confirming before booking is natural and the tool fires at the end. Our flow inverts that: identity verification is a gate — nothing can proceed until the tool returns. "Speak first, act second" costs a turn every time, and in a scripted evaluation it fails the turn outright.

Questions

  1. Is confirm-before-acting intended? The training description mentions calling "the right tools at the right time," which we read as the opposite of this.
  2. Is there a prompt pattern that makes it act on the same turn? We could not find one.
  3. Would tool_choice: "required" work here, or does that fight the fine-tune?

Happy to share more traces. Serving on Modal AutoEndpoints, vLLM, bf16, 65K context.

Sign up or log in to comment