# Preprint shows training-free method improves agents' handling of object synonyms

_Published Friday, November 6, 2026 at 2:06 AM EST · AI, Science · Latest · Tier 2 — Notable_

Researchers report in a preprint that BINDIntent, a method requiring no additional training or language-model calls, improves small language-model agents' success when instructions use synonyms for objects. On their ALFWorld-Para benchmark, synonym-paraphrase success rose by 17.5-49.1 percentage points, with 0.4 ms of CPU overhead per step.

The method matches referring expressions to observed object labels using semantic proximity and estimates of unseen alternatives. It raised GiGPO-7B's overall success rate from 61.8% to 77.4%, surpassing a prompted 24B model. The findings concern benchmark tasks under partial observability, where agents cannot see the entire environment.

## Sources

- [cs.LG updates on arXiv.org](https://arxiv.org/abs/2611.07599)

---
Canonical: https://techandbusiness.org/newswire/nQDUavueoBzgqBuzZ_biqk
Published: 2026-11-06T07:06:04.312Z
Story chronology: 2026-11-06T05:00:00.000Z
Retrieved: 2026-11-06T23:13:46.190Z
Publisher: Tech & Business (techandbusiness.org)