Encoded but Not Decoded: Layer-Localized Evidence for a Three-Level Gap in LLM Syntax
Zhenyan Lu, He Wang, Xiaohui Huang
cs.CL
Quick verdict
Claims models encode syntax they fail to deploy; careful, honest three-level design, but 48 items and verb leakage inflate small-model probe ceilings.
Summary
The paper asks whether language models fail syntactic tests because they never encode a structure or because they encode it but fail to use it at output. The authors build a 48-item trilingual (English, Chinese, German) benchmark of object- and subject-control minimal pairs and measure three quantities on identical binary decisions: behavioral deployment, LM-head (logit-lens) readout of intermediate states, and linear-probe recoverability. Across seven open-weight models (Qwen3 0.6B/14B, Llama-3.2 1B/3B, Gemma-4 E4B), the ordering behavior ≤ LM-head ≤ probe holds in the aggregate; the largest gap is 0.653 (Qwen3-0.6B Instruct, QA, subject-control: 0.250 behavior vs. 0.903 probe). The gap concentrates in subject-control, where a nearest-noun heuristic fails, persists at 14B (probe 1.000 vs. behavior 0.583), and instruction tuning degrades deployment (−14.1%) more than encoding (−11.2%). Activation patching localizes the deficit: in instruction-tuned models the causally effective layer sits about ten layers later than the probe layer. A commendably frank verb-out cross-validation shows that six of seven models' subject-control probe ceilings depend on verb identity (lexical leakage); only Qwen3-14B shows verb-invariant encoding. Main limitations: only 48 items with per-cell CIs of 0.23–0.42, LM-head readout's interpretation depends on an architectural precondition, and mechanistic evidence covers only two Qwen3 models.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- On identical binary control-dependency decisions, probe recoverability ≥ LM-head readout ≥ behavioral accuracy across seven models and three languages in the aggregate, with the surplus never negative over 14 (model, task) conditions.
- The encoding–deployment gap concentrates in subject-control (where a nearest-noun heuristic fails), persists at 14B (probe 1.000 vs. behavior 0.583 on paraphrase), and instruction tuning degrades deployment (−14.1%) more than encoding (−11.2%).
- Activation patching shows the gap is layer-localized: in instruction-tuned models the causally effective readout layer sits about ten layers later than the probe-decoded layer.
- A verb-out cross-validation shows subject-control probe ceilings in six of seven models depend on verb identity; only Qwen3-14B exhibits verb-invariant relational encoding.
- The behavior–probe gap measures the strength of shortcut-preferring decoding at the output interface.
Strengths
Measurement-comparable three-level design
All three levels (behavioral deployment, LM-head readout, probe recoverability) score the same 48 items under the same binary two-candidate decision (§3.2), so level differences are not artifacts of task or label format; the two behavioral formats (QA, paraphrase; Table 2) guard against single-template effects.
Exemplary self-scrutiny on probe validity
Appendix H's verb-out cross-validation shows six of seven models' subject-control probe peaks are leakage-dominant (Table 9 gaps +0.417 to +0.667), while only Qwen3-14B is clean (+0.014 at L22 in bv mode); the authors surface this against their own headline numbers rather than hiding it.
Systematic alternative-explanation controls
Causal evidence beyond correlation
Cross-family replication with mechanism contrast
Weaknesses
Tiny benchmark with wide uncertainty
LM-head middle tier rests on an unmet architectural precondition
Headline probe numbers are leakage-contaminated for small models
Thin family coverage for sweeping claims
Cross-lingual annotation asymmetry
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper proposes a three-level evaluation framework for syntactic knowledge in LLMs — behavioral deployment, LM-head readout (logit-lens projection through the model's own unembedding), and linear-probe recoverability — applied to identical binary decisions over a 48-item trilingual (English, Chinese, German) benchmark of object-control vs. subject-control minimal pairs (§3.1). Seven open-weight models are studied: Qwen3-0.6B base/instruct, Qwen3-14B Instruct, Llama-3.2-1B base/instruct, Gemma-4 E4B base/instruct, plus Qwen3-1.7B and Llama-3.2-3B for surplus-compression analysis. Central claims: (1) the ordering behavior ≤ LM-head ≤ probe holds across all models and languages in the aggregate, with a never-negative surplus over 14 (model, task) conditions; (2) the gap concentrates in subject-control, where a nearest-noun heuristic fails, and survives scaling to 14B (probe 1.000 vs. behavior 0.583 on paraphrase subject-control); (3) instruction tuning degrades deployment (−14.1%) more than encoding (−11.2%); (4) activation patching localizes the deficit, with the causally effective layer ~10 layers later under instruction tuning; and (5) the gap "measures the strength of shortcut-preferring decoding." Claim (5) is interpretive; the design supports it as consistency, not proof.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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
What does 'encoded but not decoded' mean in this paper?
It names the finding that an LLM's hidden states often contain syntactic information — a linear probe can recover the correct controller of a control verb (e.g., who leaves in 'John promised Mary to leave') — even when the model's own behavioral answer is wrong. The information exists internally but is not used at the output; the paper measures this with a three-level framework on identical items.
How big is the benchmark and which models were tested?
48 hand-written minimal pairs across English, Chinese, and German — half object-control ('told' type), half subject-control ('promised' type) — evaluated under question-answering and paraphrase prompts on seven open-weight models: Qwen3-0.6B base/instruct, Qwen3-14B Instruct, Llama-3.2-1B base/instruct, and Gemma-4 E4B base/instruct, plus Qwen3-1.7B and Llama-3.2-3B for scaling checks.
What is the main limitation of the study?
Sample size and probe validity. Only 48 items are used, with per-cell bootstrap confidence intervals 0.23–0.42 wide. Worse, the authors' own verb-out cross-validation shows that six of seven models' subject-control probe ceilings partly memorize verb identity rather than encoding the control relation; only Qwen3-14B shows clean, verb-invariant encoding.
Does instruction tuning help or hurt syntactic deployment here?
On these items it hurt deployment more than encoding: Qwen3-0.6B's behavioral accuracy dropped 14.1% while the probe ceiling dropped only 11.2%, and activation patching showed the layer at which interventions causally restore correct answers moved about ten layers later in instruction-tuned models.
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 1–9 are only partially recoverable from the extracted text; numeric judgments rely on in-text tables and prose restatements.
- The stated GitHub repository (items/code) could not be inspected from the text alone; availability is as claimed by the authors.
- Fresh arXiv v1 preprint (24 Sep 2026); no peer review has occurred.
- Table 6/Figure 6 extraction order is degraded; option-ordering numbers were cross-checked against prose where possible.
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.