Source-linked AI summary
VerNav: Verifier-First Low-Latency Vision-and-Language Navigation
Zhixin Wang, Chengzheyi Yao, Leyuan Liu, Xiaosong Zhang, Yongzhao Zhang
TL;DR
LLM-based VLN benefits from explicit reasoning but suffers accumulated decision-stage latency from autoregressive generation at every step. VerNav replaces that path with batched action verification, selectively invokes generation for uncertain decisions, and aligns the verifier through VPO and step-level reinforcement fine-tuning; on R2R, its verifier-only path delivers competitive navigation performance with more than 10× lower decision-stage latency.
Problem
Autoregressive reasoning at every VLN step creates substantial decision-stage latency, motivating a low-latency path that requests generation only when additional semantic cues are needed.
Method
VerNav uses batched candidate verification by default, entropy-triggered adaptive generation for compact state evidence, and two-stage alignment with VPO plus step-level reinforcement fine-tuning.
Results
VerNav achieves competitive navigation performance among representative LLM-based VLN agents while reducing decision-stage LLM latency by more than 10× on R2R.
Takeaways & Limitations
Verifier-first action selection provides a low-latency decision path while retaining adaptive semantic support for uncertain navigation decisions.
Abstract
from arXiv · showhide
Vision-and-Language Navigation (VLN) requires an agent to navigate through unseen 3D environments according to natural-language instructions. Explicit reasoning can improve instruction understanding and semantic grounding, but autoregressive generation at every step accumulates large decision-stage latency over multi-step navigation. We propose VerNav, a verifier-first framework for low-latency LLM-based VLN. The verifier reduces decision-stage latency by replacing per-step autoregressive generation with batched action verification, while an entropy-based adaptive generator is invoked only for uncertain decisions to produce compact state evidence. To further improve navigation performance with the verifier, we introduce a two-stage alignment scheme: (i) VPO improves local action-preference alignment in static verifier training, and (ii) step-level reinforcement fine-tuning provides dense progress rewards over multi-step navigation rollouts during dynamic task execution. Experiments on the Room-to-Room (R2R) benchmark show that the verifier-only decision path of VerNav achieves competitive navigation performance among representative LLM-based VLN agents while reducing average decision-stage LLM latency per step by more than $10\times$ compared with autoregressive methods.
Introduction
VerNav addresses the latency of autoregressive LLM decision-making in VLN with verifier-first action selection, selective generation, and two-stage verifier alignment. On R2R, it targets competitive navigation performance with substantially lower decision-stage latency.
- Verifier-First Action Interface: VerNav replaces per-step autoregressive generation with batched verification over executable candidate actions as the default decision path.This verifier-first interface is designed to reduce decision-stage latency.
- Entropy-Based Collaboration: Entropy routes low-uncertainty decisions through the verifier alone and invokes an adaptive generator for compact state evidence on uncertain decisions.The generator is triggered by poorly separated verifier-score distributions before verifier re-scoring.
- Two-Stage Verifier Training: VerNav combines VPO for static local action-preference alignment with step-level reinforcement fine-tuning for dynamic navigation rollouts.The two stages address both candidate ranking at individual steps and trajectory-level optimization with dense progress rewards.
- Evaluation: More than 10× lower decision-stage LLM latency is reported alongside competitive navigation performance among representative LLM-based VLN agents on R2R.The headline evaluation concerns VerNav's verifier-only decision path.
Related Work
LLM-based VLN work explores different ways to support navigation decisions through language, reasoning, maps, deliberation, and selective computation. VerNav is situated within this line of research as a verifier-first alternative focused on low-latency action selection.
- LLM-Based VLN Decision Interfaces: NavGPT generates scene descriptions and navigation reasoning, while DiscussNav coordinates multi-agent deliberation.These methods represent distinct language-based approaches to navigation decision support.
- LLM-Based VLN Decision Interfaces: MapGPT maintains a language-formalized topological map, whereas NavCoT and LangNav use task-specific training or chain-of-thought reasoning.The cited methods adapt language models to navigation through different representations or training strategies.
- Selective Reasoning and Adaptive Collaboration: Recent methods also study selective triggering and adaptive collaboration to reduce auxiliary computation while retaining semantic support when needed.Examples include switching to slower reasoning, confidence-based planning, and entropy-based triggering.
Method
VerNav frames VLN action selection as batched verification, using entropy-triggered evidence generation for uncertain states and two-stage training to align verifier preferences with navigation.
- Problem formulation: At each timestep, the agent selects movement candidates or stop from panoramic observations conditioned on the instruction and navigation history.Success requires stopping within 3 meters of the target location.
- Verifier-first interface: VerNav evaluates executable candidate actions through batched verifier queries and selects the action with the highest verification score.Each query shares the instruction, history, and candidate descriptions, differing only in the queried action; a higher “Yes” logit indicates stronger preference.
- Verifier-first interface: 0.05–0.08 seconds per step and more than 10× acceleration characterize the Raw Verifier, which nevertheless achieves only 0.09–0.38% SR on R2R val-unseen.This exposes a latency-performance gap motivating additional reasoning and alignment mechanisms.
- Reasoning-token analysis: Step-wise CoT outputs devote 47.36% of visible tokens to input restatement, 29.39% to decision making, 13.91% to state updates, and 9.32% to generic formatting.The distribution motivates assigning different functions to different modules instead of generating every role at every step.
- Entropy-based collaboration: Entropy triggers adaptive state-evidence generation when verifier scores are poorly separated, while low-entropy decisions remain on the verifier-only path.Analysis finds higher entropy for similar candidate descriptions and verifier decisions that differ from the teacher action.
- Two-stage verifier training: VPO aligns local action preferences from chosen-rejected pairs, then reinforcement fine-tuning on policy-induced rollouts supplies dense feedback for long-horizon navigation.Step-level rewards use progress toward the target, newly observed landmarks, and penalties for revisiting viewpoints, and are unavailable to the verifier at inference.
Experiments
Experiments show that VerNav combines a low-latency verifier-first interface with aligned verifier training and selective entropy-triggered evidence generation. On R2R, it maintains competitive navigation performance while substantially reducing decision-stage latency, and adaptive evidence is concentrated on uncertain trajectories and improves outcomes.
- Verifier-Side Alignment: The VPO+RFT verifier achieves the best OSR and SR on val-seen and reaches 39.63 SR on val-unseen, remaining competitive with representative LLM-based baselines.VPO raises SR from 0.39 to 38.20 on val-seen and from 0.38 to 37.25 on val-unseen; RFT further raises SR to 43.10 and 39.63.
- Decision-Stage Efficiency: 0.08s per step and 39.63 SR: VerNav’s verifier-only path reduces decision-stage latency by 12.3× relative to NavCoT with similar SR on R2R val-unseen.Under a shared Qwen2.5-3B backbone, verifier-only scoring takes 0.080s versus 2.60–26.61s for generation-based methods.
- Verifier-Side Alignment: VPO provides local chosen-rejected action-ranking supervision, while step-level reinforcement fine-tuning improves rollout-level target-reaching behavior.RFT increases OSR to 53.48 on val-seen and 51.26 on val-unseen, with a small SPL decrease on val-unseen.
- Adaptive Generation Analysis: Most trajectories require only a small number of generator calls, with trigger rates increasing at later navigation steps as verifier uncertainty accumulates.State evidence is therefore allocated selectively rather than refreshed uniformly across trajectories.
- Adaptive Generation Analysis: Failed trajectories exhibit higher entropy and more generator calls per episode than successful trajectories, concentrating adaptive generation on verifier-uncertain outcomes.The analysis groups episodes by verifier-only outcome to examine this difficulty-conditioned invocation pattern.
- Adaptive Generation Analysis: 6.0–8.0 percentage points: adding state evidence improves SR across entropy thresholds while reducing candidate entropy, and entropy-triggered invocation outperforms random and fixed-interval triggering in SR and SPL.The triggering-policy comparison uses comparable generator invocation rates.
Conclusion
VerNav addresses high decision-stage latency in LLM-based VLN by combining batched candidate verification with entropy-triggered evidence and two-stage verifier alignment. On R2R, its verifier-only path achieves competitive performance with reduced latency, while adaptive evidence selectively reduces uncertainty on triggered states.
- Conclusion: VerNav replaces per-step autoregressive action generation with batched candidate verification and uses an entropy-triggered generator only for uncertain decisions.The framework requests compact state evidence before verifier re-scoring when verifier uncertainty is high.
- Conclusion: VPO and step-level reinforcement fine-tuning align fast verification scores with local action preferences and multi-step navigation behavior.The two stages address both action-level scoring and trajectory-level navigation objectives.
- Conclusion: On R2R, the verifier-only decision path achieves competitive LLM-based navigation performance while reducing decision-stage LLM latency.Adaptive evidence is invoked selectively and reduces verifier uncertainty on triggered states.