framework
active
framework:explaining-away-concepts-shaping-the-learning-signalExplaining Away Concepts (shaping the learning signal)
Unifying framework: remove or amplify a concept's term in the exponential-tilt reward.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (4)
method
- Activation SteeringimplementsCausal intervention technique: edit NLA explanation, reconstruct via AR, use difference as steering vector to manipulate model behavior.
- Reward ShapingimplementsSubtract λ times concept score from reward.
- Data Filtering (example/token level)implementsRemove or downweight units whose preference is explained by a concept.
- Inoculation PromptingimplementsPrompt explains concept so it is not attributed to the global preference signal.
Concepts (1)
concept
- Optimal policy proportional to base policy times exp(reward/β).
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.
- High-level signals that exploit agential nature of cells to achieve outcomes without micromanaging hardware; evolution searches this space.
- Acquisition of new concepts by Bayesian model expansion and reduction.
- The biological inspiration for the paper's introspective signal; pain encodes internal evaluations guiding agents through environments
- Reports phase-like breakpoints and geometry changes as context scales; UCCT provides measurable predictor
- Training that suppresses representations of undesirable concepts may incentivize new representations, removing the detection signalhypothesis0.749Tension for monitorability.
- The pain-belief component serves as a dynamic intrinsic motivation for exploration beyond external reward
- Future direction.
- Discussion and App. A.