NeurIPS paper: visual in-context learning lifts VLM accuracy by up to 27%
cindy_x_wu · x · 2026-10-14
A NeurIPS 2026 paper argues that when vision-language models (VLMs) fail on images violating training regularities, the problem is often misleading learned priors rather than missing visual capability.
Key findings
- Controlled visual in-context learning (showing a few demonstrations) redirects the model to visual evidence and frequently corrects such errors;
- Across five vision-centric tasks, demonstrations yield large gains on prior-conflicting counterfactual examples: +27.2% for Qwen3-VL 235B and +15.1% for Qwen3-VL 32B, with prior-aligned accuracy nearly unchanged;
- Results generalize across five VLM families, with gains from +9.7% to +30.3%;
- Representation analysis shows demonstrations redirect attention toward grounding cues; ICL fails when the underlying capability is genuinely absent.
The authors conclude current VLMs possess visual grounding abilities that standard evaluations overlook.
More from Models
- Google AI Pro and Ultra subscribers get $10-$100 monthly Cloud credits to try Argon — jocarrasqueira · 2026-10-14
- Apex Matches Frontier Scores at Lower Cost but Overthinks, Says Aiden Bai — aidenybai · 2026-10-14
- Gemini 4 Argon Early Consensus: Frontier Contender With Lowest Hallucinations, But Access Was Bottleneck — johnseach · 2026-10-14
- Venice Offers Abliterated Large V2, an Unrestricted Reasoning Model for Red Teams — 0xAllen_ · 2026-10-14
- Argon 4 locked to Ultra tier, and users are balking at another $100 subscription — AIandDesign · 2026-10-14
- As Gemini 4 launches, a 'Model Launch Apology Form' roasts everyone who doubted Google — omooretweets · 2026-10-14