AI

Beyond Plausible Prediction: Safety Sufficiency in Web World Models

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/11/2026, 05:03:24 UTC.

Briefings

A new benchmark, SafeWebWorld, evaluates safety sufficiency in web world models by testing whether predicted futures support correct safety decisions. A matched human evaluation of 3,000 predictions from ten frozen open models found that while 86.1% are semantically plausible, many models show asymmetric authorization behavior, preferring BLOCK over clarifiable ASK cases. This study introduces Latent Authorisation Recovery (LAR) to assess whether evidence preservation improves safety decisions, revealing that while frozen-score recalibration improves recovery, paired recovery offers only modest gains and does not guarantee robust fine-grained authorisation.

Key points

• A matched human evaluation of 3,000 predictions from ten frozen open models found 86.1% are semantically plausible, yet models exhibit asymmetric authorization behavior preferring BLOCK over ASK.
• Latent Authorisation Recovery (LAR) introduced in SafeWebWorld benchmark shows frozen-score recalibration improves recovery but paired recovery offers only modest gains without guaranteeing robust fine-grained authorisation.

Context

Web world models increasingly predict consequences before agents act, but plausible predictions do not always support correct safety decisions. This study introduces SafeWebWorld to evaluate safety sufficiency across five web-safety domains using action-conditioned futures and policy-decisive evidence.

Limitations

Evidence is limited to a single preprint (arXiv:2610.40798v1) with no additional background or external validation. Results are based on a snapshot of 3,000 predictions from ten frozen models, and findings on recovery gains are conditional on specific fine-tuning protocols and model architectures.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 02/11/2026, 06:13:22 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeic3gkhgbvirhxplnrt4wik5ezpgd7kfocaqqq5ixrv6ahwltpdyl4