Training LLMs to know when to speak: stable partial translations fine-tuned on Qwen3-8B
burkov · x · 2026-10-30
Real-time translation must trade speed against accuracy: translating early cuts latency, but incomplete speech deprives the system of context. Standard systems rely on fixed timing rules that can't adapt to how uncertain a phrase is.
The article introduces a training method for cascaded translation (recognized speech fed to a separate translation model):
- Build "stable" partial translations by comparing translations of progressively longer source prefixes against the full-sentence translation, keeping only target text that agrees with it
- Use these examples to fine-tune Qwen3-8B so the model learns to selectively decide what is safe to emit as input arrives
The key idea: use full-sentence translations as a quality anchor to auto-generate supervision signals for reliable low-latency simultaneous interpretation.
More from Research
- ETH Zurich and NVIDIA teach a quadruped robot to solve a Rubik's cube with its feet — lukas_m_ziegler · 2026-10-30
- Long-Horizon Agents is the fastest-rising task on Papers with Code: 49 papers in 30 days — NielsRogge · 2026-10-30
- OpenFlyV2: a 1B onboard VLA model for real-time aerial drone navigation — IPEC-COMMUNITY · 2026-10-30
- Google paper: pre-reply claim checklists cut agents' sycophantic rewrites from 12.2% to 0.02% — rohanpaul_ai · 2026-10-30
- Prime gap bound shrinks from 70M to 212 since Zhang's breakthrough; twin primes needs 2 — IgorCarron · 2026-10-30
- Researcher: concept neurons alone may not be enough to induce an AI persona — Sauers_ · 2026-10-30