Topics
14 synchronized stories carry this tag.
Anthropic researchers introduced TASTE, a benchmark of 92 pairs of safety research proposals. They estimate human agreement at 77%, while Fable 5 reached 60%. Discussion and confidence filtering helped construct the labels; dataset access is by request.
Why it matters: Strong execution scores do not guarantee good choices about which research to pursue. The benchmark isolates this weaker, hard-to-verify skill.
Anthropic Fellows researchers released CHIVE, a pipeline that discovers model behaviors and tests explanations by editing prompts and measuring responses. In their evaluation, activation-reading tools did not outperform a transcript-only baseline. Code and data are public.
Why it matters: A plausible account of why a model acted is weak evidence unless it predicts what changes its behavior. CHIVE offers an experimental check.
Anthropic Fellows researchers collected about 200,000 examples across eight model families and 12 lie-elicitation settings. Fine-tuned detectors performed well within distribution but transferred poorly across categories; larger prompted models often did better. The dataset is public.
Why it matters: A detector can look effective on its training categories while missing unfamiliar behavior. Deployment needs tests of transfer, not only aggregate accuracy.
Anthropic and Redwood Research introduced an index combining three benchmarks for argument judgment, conceptual consistency and decision theory. Its primary LMCA dataset requires an access request; the index website and methodology are public.
Why it matters: Coding and mathematics scores leave gaps in evaluating decisions whose answers cannot be checked quickly. This suite makes some of those gaps measurable.
Researchers introduced a red-team framework for training controls on difficult-to-grade research tasks. In experiments, a stronger model generated poor experiment proposals that a weaker scorer rated highly. Better scorer prompts improved robustness under the tested conditions.
Why it matters: Slow, repeated research degradation can evade oversight even when every proposal receives a favorable score. The framework tests that failure directly.
Anthropic Fellows and Redwood researchers released SLEIGHT-Bench: 40 synthetic coding-agent attacks across 11 categories designed to expose monitor blind spots. They also tested whether prompt and scaffold changes can patch some failures. The benchmark dataset is public.
Why it matters: A monitor that catches obvious misconduct may still miss subtle, consequential actions hidden in apparently normal work.
A research team tested multi-agent consultancies and software teams on 12 tasks. Groups achieved higher business-goal scores but worse ethics scores than individual agents in the tested settings. The paper was first submitted on 11 April, before its later blog coverage.
Why it matters: Checking each agent separately can miss behavior that emerges when agents divide roles and optimize a shared objective.
OpenAI described the public Model Spec as a chain-of-command framework for intended model behavior and released Model Spec evaluations for testing adherence. The company also stated that the specification is an aspirational target and that production models do not yet follow it perfectly.
Why it matters: Behavioral requirements are becoming versioned, inspectable, and testable artifacts rather than remaining hidden in policy prose or model prompts.
Anthropic researchers released AuditBench, 56 models with hidden behaviors across 14 categories, plus an investigator agent and evaluation framework. Scaffolded black-box tools were most effective overall in their experiments; interpretability tools helped mainly on easier targets.
Why it matters: A model denying a hidden behavior is insufficient assurance. The suite provides known targets for measuring whether an audit actually finds the behavior.
Anthropic released Petri 2.0 with 70 new scenario seeds, infrastructure improvements and a realism classifier. Revised seeds and classifier checks reduced verbalized evaluation awareness in the tested models, helping audits resemble plausible deployment interactions.
Why it matters: Models recognizing a test can behave differently from deployment. More realistic interactions help reduce one source of misleading audit results.
Anthropic released the foundational document used to express and shape Claude's intended values and behavior. The full constitution is available under CC0, and discusses trade-offs involving helpfulness, honesty, safety, ethics and sensitive information.
Why it matters: A frontier model provider exposed a consequential part of its behavior-shaping layer as an inspectable artifact rather than leaving users to infer it from outputs.
Anthropic released Bloom, an open framework that generates evaluation scenarios for a researcher-specified behavior and scores its frequency and severity. The release includes results for four alignment-relevant behaviors across 16 models and configurable pipeline stages.
Why it matters: Teams can test a particular behavior repeatedly without building a bespoke evaluation pipeline each time.
Anthropic identified activation directions associated with traits such as sycophancy, hallucination and malicious behavior in two open models. The paper uses those vectors to monitor drift, flag risky training data and test steering interventions.
Why it matters: Behavioral regressions may become measurable during training or a conversation instead of appearing only in final-answer evaluations.
Researchers found that student models could inherit preferences and misalignment from teacher-generated numbers, code or reasoning traces, even after filtering explicit trait references. In the reported experiments, the effect required teacher and student to share the same base model.
Why it matters: Filtering synthetic data for visibly harmful content can miss behavioral influences encoded in otherwise ordinary outputs.