AI
ConflictVLA-Bench: Benchmarking Behavioral Responses of Vision-Language-Action Models to Premise Conflicts
Publication: cs.AI updates on arXiv.org
Author: Liyu Hou, Yuan Wu, Yi Chang
Source published: 29/09/2026, 04:00:00 UTC
Briefing created: 29/09/2026, 05:32:32 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 05:18:24 UTC.
Briefings
ConflictVLA-Bench introduces a benchmark to study 'Failed Persistence' in Vision-Language-Action models, where models fail on premise-consistent tasks but continue pursuing original targets. Built on LIBERO, the benchmark contains 2,826 prompt-conditioned tasks across four conflict families and two prompt conditions. Across eight VLA models, invalid premises reduced goal completion by at least 17.3 percentage points, with OpenVLA showing a 56.2 percentage point reduction. Crucially, models often retain early trajectory structure and limited action magnitude suppression even after failing, indicating terminal failure alone does not establish behavioral disengagement. The benchmark explicitly pairs conflict rollouts with premise-consistent reference rollouts to evaluate both outcomes and execution processes.
Key points
- ConflictVLA-Bench pairs conflict rollouts with premise-consistent reference rollouts to evaluate both outcomes and execution processes.
- Across eight VLA models, invalid premises reduced goal completion by at least 17.3 percentage points, with OpenVLA showing a 56.2 percentage point reduction.
Context and limitations
The benchmark covers only eight Vision-Language-Action models and eight prompt conditions across four conflict families and four structural configurations, limiting generalizability.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 29/09/2026, 05:32:32 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeifcc6wgnmsgo7rl64zit6pk2ewkqwu3nmoldvtysbx73rpghsy65u