Skip to content

v0.1.12

v0.1.12: chore: release 0.1.12
@PCfVW PCfVW tagged this 28 May 08:57 
v0.1.12 ships the Maar et al. (2026) "What's the plan?" contrastive
activation steering replication as the COLM 2026 rebuttal artefact
for the Maar-citation in Reviewers L1Vb02 and UvuC13.

Headline (full details in
docs/experiments/maar-replication/findings.md):

- Llama 3.2 3B + Maar's exact protocol + Maar's verbatim prompts
  REPRODUCES Maar's published "smaller-models" claim: baseline 60%
  → steered 30% at the documented (L = 0.8 × n_layers, m = 1.5)
  cell; all 6 binary hit-status flips are HIT→MISS, zero MISS→HIT.

- Cross-architecture: Llama 3.2 1B + Gemma 2 2B at the same
  documented cell show qualitatively different effect-directions
  (Llama inhibits, Gemma enhances).  Strength-sweep surface
  confirms this is NOT a perturbation-magnitude artefact (H3
  rejected): the contrastive direction's L2 norm ||d|| varies 10×
  across architectures (4.01 / 11.47 / 116.09), and even at
  effective-perturbation-matched strengths the qualitative
  curve shape differs (Llama monotonic-inhibition vs Gemma
  non-monotonic-enhancement with peak at m=1.0).

- Methodological implication: Maar's global m=1.5 recipe is not
  cross-architecture-transferable; their published 23-model
  rhyming-rate table is contaminated by per-architecture ||d||
  scaling.  Mechanistic (Marr Level-2/3) methods that operate on
  identified CLT features at specific positions are necessary to
  answer "where/when does the model commit?" — Level-1 behavioural
  protocols are necessary discovery tools but insufficient for
  intra-planning explanation.

Cargo.toml + Cargo.lock bumped 0.1.11 → 0.1.12.  CHANGELOG.md
[Unreleased] section moved to [0.1.12] - 2026-05-28 with a fresh
empty [Unreleased] header.  README.md version note bumped and
Paper Replications table extended with the Maar entry pointing at
the maar-replication findings.

Cargo.lock also contains an incidental windows-sys 0.61.2 → 0.59.0
downgrade picked up by cargo's dependency resolver during the v0.1.12
work — no source impact, no API change.

Pre-commit gate green:
  - cargo fmt --check (clean)
  - cargo clippy --features transformer --lib -- -W clippy::pedantic (clean)
  - cargo test --features transformer --lib  (121 passed, 0 failed)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Assets 2
Loading