Planning or Improvisation? Stress-Testing the Poetry Planning Site on Open Models and Open Cross-Layer Transcoders
cs.CL
Quick verdict
Stress test finds rhyme-plan position specificity replicates on open models only at the emission-adjacent token; the newline planning site is not recovered.
Summary
The paper stress-tests whether Lindsey et al. (2025)'s finding that Claude 3.5 Haiku plans rhymes at the newline token generalizes beyond proprietary models. On one consumer GPU, four open base models (0.6B-2.6B: Gemma 2 2B, Llama 3.2 1B, Qwen3-0.6B/1.7B) are crossed with six open cross-layer transcoders (CLTs), and the original claim is decomposed into position specificity (C1), newline site identity (C2), and a newline-resident plan (C3). Features are found bottom-up from decoder vectors since attribution graphs are unavailable. C1 replicates in 7/7 cells and 247/247 detectable prompt-by-inject pairs, but the effective position is always the final prompt token adjacent to emission, with behavioral effect sizes in only two cells (P=0.48, 0.85). C2 and C3 fail every probe: newlines are inert for the rhyme target across sweeps, a census of all active features finds no rhyme-anticipating enrichment, composition-horizon steering (36 runs, 8,640 lines) inserts the injected word at line start (703/720) without touching the rhyme six words later, and transcoder-free activation patching moves the rhyme in only 11/1,260 lines vs 4 at baseline (resolution 1.4%). Random controls show the spike location is generic and only magnitude is feature-specific. The author reads this as a scale/CLT-quality boundary condition, not a refutation. Main limitations: one prompt per cell in the grids and census, bottom-up discovery bias, and a single-model patching test.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The single-position intervention signature (C1) generalizes across all seven model-CLT cells and 247/247 detectable pairs, but the effective position is always the final prompt token adjacent to emission, not the newline.
- The newline site (C2) and a newline-resident plan (C3) are not recovered by any probe: feature census, position sweeps, composition-horizon steering over 8,640 lines, and transcoder-free activation patching all find the newline inert for the rhyme target.
- Only two of seven cells reach behaviorally meaningful steering probabilities, and effect magnitude varies six orders of magnitude across CLTs at fixed protocol, making magnitude a transcoder property rather than a model property.
- Random write-directions concentrate their leverage at the same final-token position, so the spike location is generic and only its magnitude distinguishes rhyme features.
- The overall result is a boundary condition on scale and transcoder quality rather than a refutation of Lindsey et al. (2025); Figure 13's shape is reproduced, not its mechanism.
Strengths
Rigorous claim decomposition with explicit epistemic scope
The original claim is split into C1/C2/C3 (Section 2) with a protocol-difference matrix (Table 1) stating which conclusions each probe can and cannot support; the word 'refute' is deliberately never used, and 'not recovered' is defined precisely.
Unusually deep control suite
Random features and Gaussian directions separate spike location (generic, 10/10 Gemma draws) from magnitude (feature-specific, 2-5 orders larger); random-weight and weight-shuffled models abolish the spike (9,974,880x vs 3.5x/99x, Section 5); bit-exact self-patch no-ops validate the patching harness (Section 4.5); an MLP-reconstruction check exposed the authors' own encoder-hook bug (Appendix D).
Power-aware negative results
Transcoder-free patching closes the 'wrong features' loophole
Transparent handling of its own metric confound
Weaknesses
Single-prompt design in headline experiments
Untestable scale gap
The decisive test is single-model
Breadth re-analysis inherits earlier defective code
Best-of-grid headline numbers
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Did this paper reproduce Anthropic's rhyme-planning finding on open models?
Partially. It reproduced the causal shape of the original Figure 13 — an intervention that works at exactly one position — but on Gemma 2 2B, Llama 3.2 1B, and Qwen3 models that position is always the final prompt token right before the rhyming word, not the newline token where Anthropic located rhyme planning. No probe (feature census, steering, whole-residual patching) found any rhyme plan at the newline, so the mechanism did not replicate at 0.6B-2.6B scale.
Why might the newline planning site fail to appear in smaller models?
The paper offers three readings: (i) scale — planning may only emerge above roughly 2.6B parameters, consistent with prior work showing planning signals strengthen with model size; (ii) the plan could be represented in a form invisible to open transcoders — weakened by the fact that transcoder-free activation patching also found nothing at the newline; (iii) these models genuinely improvise the rhyme at emission. The author ultimately reports 'not recovered' rather than claiming improvisation is proven.
What is a cross-layer transcoder (CLT) and why does it matter here?
A CLT is a dictionary-learning model descended from sparse autoencoders whose features read and write residual-stream information across layers. Anthropic's original result used a proprietary 30M-feature CLT on Claude 3.5 Haiku. This paper uses six openly available CLTs from two independent pipelines (mntss ReLU, BlueLightAI JumpReLU) and finds that effect magnitude varies six orders of magnitude across them at fixed protocol, so the choice of transcoder is itself a major experimental variable.
Can the experiments be rerun, and at what cost?
Yes, per the paper: code, data, and per-number committed artifacts are public on GitHub, each model-CLT cell reproduces with three commands, and everything runs on a single 16 GB consumer GPU (RTX 5060 Ti). One-time model and CLT downloads are 17-41 GB, the vocabulary scan takes 3-4 minutes, and grids take seconds per strength. Note that Gemma 2 and Llama 3.2 are gated on HuggingFace and require accepting license terms.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (position sweeps, Figure 3 controls) are available only as textual descriptions in the extracted text; log-scale axis details and bar placements could not be visually verified.
- The abstract in the extraction is truncated mid-sentence, though the corresponding content appears complete in Sections 4.5 and 7.
- Some cross-references appear as raw LaTeX labels (e.g., '§efsec:patching'), a minor extraction artifact that does not affect comprehension.
- Repository contents (Jacopin 2026b), committed artifacts, and registered-criteria files could not be inspected; reproducibility ratings rely on the paper's descriptions.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.