When Tokenization is Secretly Output Supervision
Tanja Baeumel, Josef van Genabith, Simon Ostermann
cs.CL
Quick verdict
Argues output tokenization is the true supervision signal; clean decoupling + probing at toy scale support it, but production-scale evidence stays indirect.
Summary
The paper argues that in autoregressive language models tokenization is not merely input preprocessing: because the training loss is computed over output tokens, output tokenization granularity defines what a model must resolve in a single forward pass, and thus the supervision signal itself. To test this, the authors train roughly 3.4M-parameter decoder-only transformers on three-digit addition in a 2×2 factorial design that decouples input and output tokenization through a shared 1003-token vocabulary, totalling 720 runs across seeds and hyperparameters. Training dynamics and accuracy track the output side: fragmented-output models reach about 1.00 exact match while holistic-output models plateau near 0.85 at the best setting. Linear probes support a 'minimal computation hypothesis': at the '=' position, fragmented-output models decode only the first result digit, whereas holistic-output models decode all three. A big-endian ablation shows future digits become decodable above chance when carry lookahead is task-required, separating gradient reward from task requirement. A survey of 120 recent *CL numeric-reasoning papers finds only 9.2% report their models' numeric tokenization, while 69% compare across tokenization regimes without flagging it. Main limitations: toy-scale, single-task evidence; only indirect production-scale support; no significance testing; and a survey resting on title-keyword search and author annotation without inter-annotator agreement; no code release is stated.
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
- In autoregressive training the loss is computed over output tokens, so output tokenization granularity — not input representation — defines the atomic unit of supervision and thus the learning problem.
- In the 2×2 decoupling experiment, task performance, training dynamics, and internal representations group by output tokenization regime (MF ≃ MHin,Fout; MH ≃ MFin,Hout), largely independent of input tokenization.
- The minimal computation hypothesis holds: fragmented-output models make only the currently-required result digit decodable at '=', holistic-output models make all digits decodable; under big-endian, task-forced lookahead makes future digits decodable above chance even without gradient pressure.
- Only 9.2% (11/120) of surveyed 2024–2025 *CL numeric-reasoning papers report the numeric tokenization of evaluated models, while 69% (83/120) compare across tokenization regimes without flagging it, so cross-model conclusions risk conflating task definition with reasoning ability.
Strengths
Causal decoupling design isolates the output axis
All four conditions share one 1003-token vocabulary (App. A.2), so input and output tokenization vary independently with zero architectural change; cross-entropy is masked to output tokens only (App. A.3), making the gradient signal a pure function of output tokenization. The factorial predictions (MF ≃ MHin,Fout, MH ≃ MFin,Hout) are stated before results (Sec. 3.3).
Falsifiable hypothesis tested with converging, honestly-reported probe evidence
The minimal computation hypothesis (Sec. 2.2) yields two concrete predictions confirmed in Fig. 4/App. C with 720 probes (linear, MLP, circular). Anomalies are openly discussed: above-chance MLP readings at layers 1–2 for MFin,Hout attributed to probe capacity (they vanish at layer 3), and circular-probe failure on MH attributed to non-modular digit encoding.
The big-endian ablation preempts the strongest objection
Unusually strong reproducibility detail for the training experiments
Well-scoped claims and honest survey caveat
Weaknesses
Toy-scale, single-task evidence for a scale-spanning claim
Loss masking to output tokens diverges from production pretraining
Equivalence claims without inferential statistics
Survey robustness unquantified
No stated code/data release; thin checkpoint base for probes
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a position paper backed by controlled experiments and a survey. The central claim: in autoregressive LMs the cross-entropy loss is computed over output tokens, so output tokenizer granularity — not input formatting — is the operative supervision variable that defines the learning problem. Three sub-claims structure the paper: (1) a 2×2 decoupling experiment (Sec. 3.3) shows performance and training dynamics track the output axis; (2) the 'minimal computation hypothesis' (Sec. 2.2) — models internally resolve exactly what the per-position loss demands — is confirmed by probing (Sec. 3.4); (3) a survey of 120 *CL papers (Sec. 4) shows the resulting cross-model confound goes unreported: 11/120 (9.2%) report numeric tokenization, 83/120 (69%) compare across regimes without flagging it.
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 it mean that tokenization is 'output supervision'?
In autoregressive language models the training loss is computed one output token at a time. So whether a value like 578 is emitted as one token or as three digit tokens 5,7,8 determines what the model must compute within a single forward pass. The paper argues output tokenization therefore acts as a task definition — it shapes what the model is trained to solve — rather than merely an input formatting choice.
Do the findings apply to large production LLMs like Llama or Qwen?
The authors argue the core logic is scale-independent because loss is computed over output tokens at any model size, but their experiments use only ~3.4M-parameter models on three-digit addition. Evidence at production scale is indirect: their prior work found digit-position-specific circuits in holistic-tokenizer models, but there input and output share one tokenizer, so output supervision cannot be isolated. Direct large-scale manipulation remains future work.
What is the minimal computation hypothesis?
It is the paper's claim that a model internally resolves exactly what its per-token loss requires, and feels no direct pressure to resolve more. Probing confirmed it: models with digit-level (fragmented) output tokenization made only the first result digit decodable at the '=' position, while models with whole-number (holistic) output tokenization made all three digits decodable. When the task itself forced lookahead (big-endian digit order with carries), future digits became decodable above chance anyway.
What did the survey of 120 papers find?
Across 120 ACL/EMNLP/NAACL/EACL papers from 2024–2025 on numeric or mathematical reasoning, only 11 (9.2%) reported the numeric tokenization strategy of the models they evaluated, and 83 (69% of all surveyed) compared models that use different tokenization regimes — holistic vs fragmented — without flagging it. The authors' point is not that those comparisons are wrong, but that the potential confound cannot be assessed from the published record.
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.
- All figures (Figs 2–17) are raster images in the extraction; curve-level details (per-seed variance, exact layer-wise probe values) could only be assessed via captions, in-text numbers, and Tables 1–3.
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.