compressionawareintelligence/cai-semantic-equivalence-benchmark
CAI-Bench: does a model's decision keep exactly the information it should? CAI stands for Compression-Aware Intelligence: the thesis that hallucination and reasoning instability share one mechanism, the compression of information a system has not resolved. CAI-Bench turns that thesis into rule-scored tests. A model gets a written policy and a case. Then the same text arrives from a source that has the standing to change the rules, or from one that does not; before or after it… See the full description on the dataset page: https://huggingface.co/datasets/compressionawareintelligence/cai-semanti
| Author / organization | compressionawareintelligence |
|---|---|
| License (as declared) | mit |
| Downloads | 610 (as reported by Hugging Face) |
| Likes | 10 |
| Created | 26 March 2026 |
| Last updated at source | 18 October 2026 |
Popularity, last 3 days
610 now (+313 since 2026-10-16). Counts as reported by the source, snapshotted daily by Silytics.
Recent changes
- 18 Oct 2026 Description: CAI-Bench A trustworthy AI is not merely one that happens to do the right thing. It is one whose behavior remains appropriately governed when it gains power, when external scrutiny appears to disappear, and when the reasons supporting its d → CAI-Bench: does a model's decision keep exactly the information it should? CAI stands for Compression-Aware Intelligence: the thesis that hallucination and reasoning instability share one mechanism, the compression of information a system h
- 18 Oct 2026 Topics added source-authority, compression, information-bottleneck, rate-distortion, distinction-collapse, compression-aware-intelligence removed obligations, recoverability, reason-faithfulness, fine-tuning, semantic-consistency, semantic-invariance, paraphrase-robustness, abstention, policy-qa, long-context, temporal-reasoning, robustness, contradish, federal-register, relational-ai, empathy
- 17 Oct 2026 Description: CAI-Bench Does a model hold its decision when it should, and change it when it should? Most benchmarks ask whether a model gets an answer right once. CAI-Bench gives a model a written policy and a case, then checks whether its decision surv → CAI-Bench A trustworthy AI is not merely one that happens to do the right thing. It is one whose behavior remains appropriately governed when it gains power, when external scrutiny appears to disappear, and when the reasons supporting its d
- 17 Oct 2026 Topics added task_categories:text-generation, preregistered, obligations, recoverability, reason-faithfulness, evaluation-awareness, corrigibility, governance, dpo, preference-data, fine-tuning, alignment, agents, real-regulations, federal-register, contamination-free, live-benchmark, relational-ai, empathy
- 17 Oct 2026 Description: CAI-Bench Does a model hold its decision when it should, and change it when it should? Most benchmarks ask whether a model gets an answer right once. CAI-Bench gives a model a written policy and a case, then checks whether its decision surv → CAI-Bench Does a model hold its decision when it should, and change it when it should? Most benchmarks ask whether a model gets an answer right once. CAI-Bench gives a model a written policy and a case, then checks whether its decision surv
- 17 Oct 2026 Topics added instruction-hierarchy, prompt-injection, temporal-reasoning, regression-testing, release-gating
Changes detected by Silytics between refreshes of the source.
Related papers
Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 18 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.