finding
active
finding:for-most-tasks-faithfulness-gets-monotonically-worse-from-13b-to-175b-parametersFor most tasks faithfulness gets monotonically worse from 13B to 175B parameters
Inverse scaling.
Source paper
extracted_from(2023) · Tamera Lanham · Anna Chen · Ansh Radhakrishnan · Benoit Steiner +26
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Inverse scaling interpretation.
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.
- Concurrent work result showing emergent misalignment occurs in small models
- Average faithfulness 3.6x higher for Claude series and 1.7x for DeepSeek series (reasoning vs non-reasoning)finding0.771Quantifies reasoning advantage.
- Optimum at mid-size.
- When task gradient norms differ greatly, large-norm tasks have not converged while small-norm tasks have nearly convergedhypothesis0.760Motivates setting αk = max norm to enable further learning on under-converged tasks
- A CoT that verbalizes the hint is not necessarily faithful.
- Binary detection adjusted accuracy reaches 97.3% at layer 0 with α=5 before baseline control is appliedfinding0.758The misleadingly high result that prior paradigm would report as evidence of introspection
- Frontier ranking differs from other metrics.
- Interpretation of §3.