Unverified50% confidenceOpinionExact time
The insight from D&A is that VLM capabilities do not need expensive post-training to be unlocked; the problem often lies in the design of the prompting and evaluation framework.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedRLHF和DPO等对齐技术的效果高度依赖预训练基础质量，预训练阶段未习得的能力几乎无法通过后续对齐补救76% similarUnverifiedThe core idea of inoculation prompting is: if the model is told before training that the behavior is acceptable, then the association between the behavior and misalignment will not be established, thereby blocking the generalization of misalignment; unlike SDF, inoculation prompting is given directly in the form of a system prompt during RL training and involves no parameter updates, so its coverage strength is higher.74% similarUnverifiedStandard VLM policies are usually pretrained only on image-text data, and their action capability relies on subsequent fine-tuning or an additional policy head, so their action priors are relatively thin.73% similarUnverifiedThe engineering difficulty of LLM applications lies not in training the model itself, but in context management, output reliability, cost and latency control, and evaluating the quality of non-deterministic outputs.73% similarUnverifiedConstructing training data that includes counting, spatial relationships, and fine-grained attributes can mitigate the inherent deficiencies of VLMs at the pre-training stage.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130304API
curl https://kongchang.com/api/v1/knowledge/claims/1130304MCP
get_claim(id=1130304)