concept
active
concept:harness-activation-failureHarness Activation Failure
A failure mode where weak-tier models fail to invoke relevant harness artifacts (e.g., skills) during task solving
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (3)
concept
- Harness-Benefit Capabilityassociated_withThe capability of a task-solving agent to benefit from updated harnesses during task solving
- Skill-Load Rateassociated_withThe fraction of trajectories in which an agent actively loads at least one skill into its context
- Format Gate (SkillsBench)associated_withSkillsBench enforcement mechanism that accepts only single-key JSON actions; composite multi-key actions are rejected, preventing skill loading
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- A failure mode where even when harness artifacts are loaded, weak-tier models fail to follow their guidance faithfully
- The capability of an evolver model to produce useful persistent harness updates from execution evidence
- Internal representations of the model on which probes operate; the method uses activations to rank datapoints.
- Standard method in mechanistic interpretability that intervenes on activations; VPD flips this paradigm by patching parameters.
- The external non-parametric context and infrastructure (prompts, skills, memories, tools) through which an LLM is deployed for task execution
- Natural-language harness artifacts that encode standing behavioral rules, task policies, and reasoning procedures
- Clamping activations along the Assistant Axis to remain above a minimum threshold (25th percentile), introduced as a stabilization method
- The conventional approach (e.g., SAEs, transcoders) of decomposing activations into interpretable features.