{"id":"eebee87e-074e-41e8-b21d-ff561f01b1df","arxiv_id":"2607.05316","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"Linear probes on frozen LLM hidden states recover an approximate remaining-output-length signal that is decodable at prompt-end, transfers across datasets, and shifts upward at retraction tokens.","lead":"This paper shows that LLMs linearly encode an estimate of how many tokens remain in their response, readable from hidden states before generation begins and updating mid-generation when the model retracts. A smart generalist might read it because it suggests LLMs carry internal 'plans' for output length, with potential applications in early termination and detecting unfaithful reasoning.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The prompt-end decodability — the paper's foundational claim — is not controlled against simple prompt-content features (e.g., prompt token length, question type) that mechanically correlate with completion length, leaving open whether the probe reads a length plan or just prompt content.","rationale":"The reader identified tokenization as an alternative explanation for the transfer asymmetry. I agree this is a real gap, but I think the more load-bearing version is broader: the paper lacks any prompt-feature control on its foundational decodability claim. Tokenization is one specific prompt feature; the general class includes prompt length, question type, difficulty, and content complexity — all of which could confound the probe's length signal. The reader's concern is a specific instance of this broader issue.\n\nThat said, my concern is addressable with a simple control (prompt-length regression baseline) that the authors could run on their existing cached data. The natural-language results may survive even after controlling for prompt length — the probe does have access to the full hidden state, which encodes richer information than token count alone. The concern is that we don't know yet whether it survives, and the 'plan-like representation' interpretation depends on it surviving.\n\nI recommend UNCHANGED rather than a harsher verdict because: (1) the paper is transparent about its limitations and explicitly frames the 'plan' interpretation as representational, not causal; (2) the within-distribution results (Table 2) show the Remaining Count Probe beating both the constant-median and exact-countdown baselines on several natural-language datasets, which is harder to explain by prompt features alone since mid-completion positions are used; (3) the concern is testable without new experiments. The CONDITIONAL verdict is appropriate — the results are promising but the 'plan-like' interpretation requires the prompt-feature control I've described before it can be fully endorsed.\n\nThe retraction-spike evidence (claim 3) is the weakest of the three claims, as the reader noted. It is qualitative, drawn from the worst-MAE region, and lacks a length-matched control. But this is already acknowledged by the paper and reflected in the CONDITIONAL verdict. My concern about prompt-content confounds is independent of the retraction evidence and affects the more fundamental claims (1) and (2).","tokens_in":17041,"tokens_out":10676,"duration_ms":468792,"concrete_test":"For each dataset, train a linear regression predicting T from prompt token count alone (and optionally prompt character count + a one-hot dataset indicator). Compare its prompt-end MAE to the Completion Length Probe's MAE in Table 1. If the prompt-length baseline approaches the probe's MAE on any natural-language dataset (within ~20%), the 'plan-like representation' interpretation weakens for that dataset. Additionally, for the cross-dataset transfer matrix (Table 3): residualize both the probe predictions and ground-truth T against prompt token length, then recompute the matrix. If the natural→synthetic transfer advantage disappears or reverses after residualization, the 'general length-tracking direction' interpretation is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's first and most fundamental claim is that total response length T is linearly decodable from the prompt's last hidden state alone (Table 1). The only baseline is the constant-median predictor, which uses zero prompt information. No baseline controls for simple prompt features that could correlate with completion length.\n\nThis confound operates at two levels. On synthetic datasets (Count/Countdown), the prompt contains the number n which mechanically determines T ≈ n+1. The probe's near-perfect performance on Countdown (MAE 5.27 for Llama) almost certainly reflects reading n from the prompt rather than maintaining a length plan. The paper acknowledges this ('verifiably reading the relevant information'), using synthetic data as a positive control — but this means synthetic results are not evidence for a plan-like representation.\n\nOn natural-language datasets, the confound is subtler but still unaddressed. Prompt features (token count, question type, difficulty indicators) correlate with response length: a multi-step GSM8K problem likely produces a longer solution than a simple one. The probe's improvement over the constant-median baseline (e.g., GSM8K: 42.29 vs 58.66) could partially or largely reflect these content correlations rather than an internal length estimate. The constant-median baseline is too weak to distinguish these because it uses no prompt information whatsoever.\n\nThis confound also provides an alternative explanation for the cross-dataset transfer asymmetry (§5). A probe trained on diverse natural-language data might learn general prompt-content-to-length correlations that transfer broadly (including to synthetic data where content features strongly predict length), while a synthetic-trained probe learns a task-specific number-reading mapping that doesn't generalize. This produces the observed one-way asymmetry without requiring a 'general length-tracking direction.'\n\nThe paper's argument against tokenization (§5, 'unparsimonious as the lo","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"The paper investigates whether LLMs maintain an internal, linearly decodable estimate of remaining output length. Using minimal-capacity linear probes on frozen hidden states of three 7-8B instruction-tuned models across seven datasets (two synthetic, five natural-language), the authors report three findings: (1) total response length T is linearly decodable from the prompt's last hidden state alone, beating a constant-median baseline in every cell; (2) probe directions trained on natural-language datasets transfer cross-dataset, including to synthetic targets, while the converse (synthetic to natural) fails—an asymmetry the authors attribute to natural-language training recovering a more general length-tracking direction; (3) on curated high-MAE completions, the probe's per-position estimate shifts upward at retraction tokens, a behavior monotonic baselines cannot reproduce. The paper is methodologically careful in several respects: three-seed averaging, a stratified probe family separating prompt-end decodability from mid-completion updates, and an explicit limitations section acknowledging that decodability does not entail causal use and that the retraction evidence is qualitative.","tokens_in":17752,"tokens_out":1324,"duration_ms":113576,"significance":"The paper addresses a well-motivated question at the intersection of mechanistic interpretability and LLM generation dynamics. Its strengths include a clean probe family design that stratifies contributions (prompt-end vs. mid-completion), a cross-dataset transfer matrix across three model families that goes beyond standard within-distribution probing, and honest scoping of claims (e.g., explicitly framing the retraction result as qualitative, acknowledging the worst-MAE selection bias). The cross-dataset asymmetry result, if it holds up under the controls discussed below, is a genuinely informative finding for the probing literature. The code is released (anonymously), and the hyperparameter and dataset configuration tables (Tables 6-7) support reproducibility. The per-layer sweep (Appendix A.1) and the classification-probe ablation (Appendix A.8) are valuable additions.","major_comments":[{"comment":"§4.1, Table 1: The paper's foundational claim is that total response length T is linearly decodable from the prompt's last hidden state alone. The only baseline is the constant-median predictor, which uses zero prompt information. No baseline controls for simple prompt features (e.g., prompt token length, question type, structural cues) that could mechanically correlate with completion length. On natural-language datasets, the probe's improvement over the constant-median baseline (e.g., GSM8K: 42.29 vs. 58.66 for Llama) could partially reflect these content correlations rather than an internal length estimate. A simple regression baseline using prompt token count (and perhaps one or two other surface features) as a predictor of T would help distinguish 'the model encodes a length plan' from 'the model encodes prompt features that correlate with length.' This is load-bearing for the paper","section":null},{"comment":"§5, Table 3: The cross-dataset transfer asymmetry (natural → synthetic transfers; synthetic → natural fails) is a central result. The paper argues that tokenization differences are 'unparsimonious as the load-bearing explanation' because a tokens-per-character story predicts symmetric failures, but the matrix shows one-way generalization. However, the argument against a tokenization confound is made verbally rather than empirically. The paper itself acknowledges that 'a shared-BPE replication is left to follow-up work.' Given that this asymmetry underpins the 'plan-like representation' interpretation, at minimum the paper should more precisely characterize what tokenization differences exist across the datasets and why the asymmetry argument is robust to them. Alternatively, a control using a fixed tokenizer across datasets would substantially strengthen the claim. Without this, an artif","section":null}],"minor_comments":[{"comment":"§3.3: The Completion Length Probe is trained with MSE loss but evaluated with MAE. The rationale for this loss/metric mismatch is not stated. A brief note on why MSE was chosen for training while MAE is the headline metric would help.","section":null},{"comment":"Table 2: Several cells show the Remaining Count Probe losing to Exact countdown (e.g., Llama/Count: 34.41 vs. 29.80; Llama/MMLU-Pro: 123.56 vs. 123.31). The paper addresses this in the 'Implied claim' paragraph, but marking these cells explicitly (e.g., with a different symbol) would improve clarity.","section":null},{"comment":"Figure 3: The table in panel (b) lists token positions 179-191, but the text references t=173 with r_t=814. The relationship between the table range and the referenced position is unclear; better alignment would help.","section":null},{"comment":"§A.9: The selection procedure for the four additional retraction examples is described as sorting by per-completion MAE and walking down from the worst. It would help to state how many completions were examined before these four were selected, to give a sense of prevalence.","section":null},{"comment":"Table 7: The Count and Countdown datasets have only 301 train and 301 eval samples. This is quite small compared to the natural-language datasets. A note on whether this affects the cross-dataset transfer results (e.g., whether the synthetic→natural failure is partly a sample-size artifact) would be appropriate.","section":null},{"comment":"§2: The related work section is well-structured but could briefly mention prior work on length-controlled generation or length prediction in NLP, as these are adjacent to the prompt-end decodability claim.","section":null}],"recommendation":"major_revision","confidential_remarks":"The reader's report and stress-test note both flag the prompt-content confound as the primary concern. On reading the paper, I agree this is the most important issue: the constant-median baseline is too weak to distinguish 'the model has a length plan' from 'the model encodes prompt features that correlate with length.' This is fixable with a relatively simple additional baseline (regression on prompt token count and a few surface features), so I recommend major revision rather than rejection. The cross-dataset transfer result is the paper's most novel contribution, but the tokenization confound is acknowledged but not empirically addressed; the authors should at least sharpen the argument or provide partial evidence. The retraction-spike claim is appropriately hedged and does not need to be load-bearing for the paper's main contribution."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive report. Both major comments identify controls that would strengthen the paper's central claims. We address each below and indicate what revisions we will make.","responses":[{"response":"This is a fair and important point. We agree that the constant-median baseline is the weakest possible control and that a prompt-feature regression baseline is needed to isolate what the probe is reading beyond surface correlations. We will add this baseline in revision. Specifically, we will train a simple linear regression (and optionally a gradient-boosted regressor) mapping a small set of prompt-level surface features — prompt token count, prompt character count, and question type indicator — to T, fit on the train split and evaluated on the eval split, and report its MAE alongside the constant-median and Completion Length Probe in Table 1. This directly tests whether the probe's improvement over the constant baseline is attributable to surface features that are linearly readable from the prompt alone. We expect the probe to outperform this baseline on the synthetic datasets (where T is a deterministic function of the prompt and the probe recovers it almost exactly, e.g., MAE 5.27 on Countdown), but the comparison on natural-language datasets is the empirically open question the referee rightly flags. We note two partial defenses of the current claims that the revision will make explicit: (1) The cross-dataset transfer result (Table 3) is difficult to explain by surface-feature correlations alone — a probe trained on OpenThoughts-1k transfers to Count/Countdown, which have a completely different surface structure, and a probe trained on MMLU-Pro transfers to GSM8K despite different question formats. If the probe were merely reading 'prompt token count' or 'question type,' it should not generalize across these structurally dissimilar datasets. (2) The per-layer sweep (Appendix A.1) shows the signal is absent at layer 0 (the embedding output) and emerges only in later","revision_made":"no","referee_comment":"§4.1, Table 1: No baseline controls for simple prompt features (prompt token length, question type, structural cues) that could mechanically correlate with completion length. A regression baseline using prompt token count would help distinguish 'the model encodes a length plan' from 'the model encodes prompt features that correlate with length.'"},{"response":"We agree that the tokenization argument should be supported empirically rather than only verbally. We will add two things in revision. First, a quantitative characterization of tokenization differences across all seven datasets: for each dataset, we will report the distribution of tokens-per-character, BPE merge depth, and vocabulary overlap with the synthetic Count/Countdown sets. This makes the tokenization landscape explicit and lets the reader assess whether the datasets that transfer well to each other are also the ones with similar tokenization profiles. Second, we will strengthen the logical argument with a concrete analysis: if tokenization mismatch were the load-bearing explanation for transfer failure, we would expect the degree of tokenization mismatch between a train-target pair to predict the transfer MAE. We will test this by computing the correlation between a tokenization-similarity metric (e.g., vocabulary overlap or tokens-per-character ratio) and the off-diagonal entries of the transfer matrix. If tokenization is the primary driver, this correlation should be high; if it is not, the correlation should be weak — which is what we expect given the asymmetry (natural→synthetic transfers despite large tokenization differences; synthetic→natural fails despite the same differences). We note that the asymmetry itself is logically inconsistent with a pure tokenization story: the tokenization gap between, say, OpenThoughts-1k and Count is the same in both directions, yet transfer succeeds in one direction and fails in the other. A tokens-per-character confound predicts symmetric failure, not asymmetric. However, we acknowledge that tokenization could interact with the probe-fitting dynamics in ways this simple symmetry argument does not capture — for instance, ","revision_made":"no","referee_comment":"§5, Table 3: The argument against a tokenization confound for the cross-dataset transfer asymmetry is made verbally rather than empirically. The paper should characterize tokenization differences across datasets and explain why the asymmetry argument is robust to them, or provide a fixed-tokenizer control."},{"response":"We agree that the tokenization argument should be supported empirically rather than only verbally. We will add two things in revision. First, a quantitative characterization of tokenization differences across all seven datasets: for each dataset, we will report the distribution of tokens-per-character, BPE merge depth, and vocabulary overlap with the synthetic Count/Countdown sets. This makes the tokenization landscape explicit and lets the reader assess whether the datasets that transfer well to each other are also the ones with similar tokenization profiles. Second, we will test whether tokenization similarity predicts transfer performance: if tokenization mismatch were the load-bearing explanation, the degree of mismatch between a train-target pair should correlate with the off-diagonal transfer MAE. We will compute this correlation and report it. We note that the asymmetry itself is logically inconsistent with a pure tokenization story: the tokenization gap between, say, OpenThoughts-1k and Count is the same in both directions, yet transfer succeeds in one direction and fails in the other. A tokens-per-character confound predicts symmetric failure, not asymmetric. However, we acknowledge that tokenization could interact with probe-fitting dynamics in ways this simple symmetry argument does not fully capture — for instance, if the synthetic-trained probe overfits to a narrow direction that happens to align with a tokenization artifact, the failure on natural-language targets could be a generalization failure rather than a tokenization mismatch per se. The correlation analysis we propose would help distinguish these. Regarding a fixed-tokenizer control: this would require regenerating all completions with a shared BPE, which is a substantial experimental effort that","revision_made":"no","referee_comment":"§5, Table 3: The argument against a tokenization confound for the cross-dataset transfer asymmetry is made verbally rather than empirically. The paper should characterize tokenization differences across datasets and explain why the asymmetry argument is robust to them, or provide a fixed-tokenizer control."},{"response":"We agree that the tokenization argument should be supported empirically rather than only verbally. We will add two analyses in revision. First, a quantitative characterization of tokenization differences across all seven datasets: for each dataset, we will report the distribution of tokens-per-character, vocabulary overlap with the synthetic Count/Countdown sets, and average BPE fragment length. This makes the tokenization landscape explicit. Second, we will test whether tokenization similarity between a train-target pair predicts the off-diagonal transfer MAE. If tokenization mismatch were the load-bearing explanation, this correlation should be high; if it is not, the correlation should be weak. We expect the latter, given the asymmetry: the tokenization gap between OpenThoughts-1k and Count is the same in both directions, yet transfer succeeds in one direction and fails in the other. A pure tokens-per-character confound predicts symmetric failure, not asymmetric. We acknowledge, however, that tokenization could interact with probe-fitting dynamics in subtler ways — for instance, if the synthetic-trained probe overfits to a narrow direction that happens to align with a tokenization artifact, the failure on natural-language targets could reflect poor generalization rather than a direct tokenization mismatch. The correlation analysis would help distinguish these explanations. Regarding a fixed-tokenizer control: this would require regenerating all completions with a shared BPE across three model families, which is a substantial experimental effort. We will scope this for follow-up work but believe the two analyses above, combined with the existing asymmetry argument, substantially narrow the space of tokenization-based explanations. We will revise the manuscript to make","revision_made":"no","referee_comment":"§5, Table 3: The argument against a tokenization confound for the cross-dataset transfer asymmetry is made verbally rather than empirically. The paper should characterize tokenization differences across datasets and explain why the asymmetry argument is robust to them, or provide a fixed-tokenizer control."},{"response":"We agree that the tokenization argument should be supported empirically rather than only verbally. We will add two analyses in revision. First, a quantitative characterization of tokenization differences across all seven datasets: for each dataset, we will report the distribution of tokens-per-character, vocabulary overlap with the synthetic Count/Countdown sets, and average BPE fragment length. This makes the tokenization landscape explicit and lets the reader assess whether datasets that transfer well to each other also share similar tokenization profiles. Second, we will test whether tokenization similarity between a train-target pair predicts the off-diagonal transfer MAE: if tokenization mismatch were the load-bearing explanation, this correlation should be high; if it is not, the correlation should be weak. We expect the latter, given the asymmetry — the tokenization gap between OpenThoughts-1k and Count is the same in both directions, yet transfer succeeds one way and fails the other. A pure tokens-per-character confound predicts symmetric failure, not asymmetric. We acknowledge, however, that tokenization could interact with probe-fitting dynamics in subtler ways: for instance, if the synthetic-trained probe overfits to a narrow direction that happens to align with a tokenization artifact, the failure on natural-language targets could reflect poor generalization rather than a direct tokenization mismatch. The correlation analysis would help distinguish these. Regarding a fixed-tokenizer control: this would require regenerating all completions with a shared BPE across three model families, which is a substantial experimental effort. We will scope this for follow-up work but believe the two analyses above, combined with the existing asymmetry argument, meaning","revision_made":"no","referee_comment":"§5, Table 3: The argument against a tokenization confound for the cross-dataset transfer asymmetry is made verbally rather than empirically. The paper should characterize tokenization differences across datasets and explain why the asymmetry argument is robust to them, or provide a fixed-tokenizer control."},{"response":"We agree that the tokenization argument should be supported empirically rather than only verbally. We will add two analyses in revision. First, a quantitative characterization of tokenization differences across all seven datasets: for each dataset, we will report the distribution of tokens-per-character, vocabulary overlap with the synthetic Count/Countdown sets, and average BPE fragment length. This makes the tokenization landscape explicit and lets the reader assess whether datasets that transfer well to each other also share similar tokenization profiles. Second, we will test whether tokenization similarity between a train-target pair predicts the off-diagonal transfer MAE: if tokenization mismatch were the load-bearing explanation, this correlation should be high; if it is not, the correlation should be weak. We expect the latter, given the asymmetry — the tokenization gap between OpenThoughts-1k and Count is the same in both directions, yet transfer succeeds one way and fails the other. A pure tokens-per-character confound predicts symmetric failure, not asymmetric. We acknowledge that tokenization could interact with probe-fitting dynamics in subtler ways: for instance, if the synthetic-trained probe overfits to a narrow direction that happens to align with a tokenization artifact, the failure on natural-language targets could reflect poor generalization rather than a direct tokenization mismatch. The correlation analysis would help distinguish these. Regarding a fixed-tokenizer control: this would require regenerating all completions with a shared BPE across three model families, which is a substantial experimental effort. We will scope this for follow-up work but believe the two analyses above, combined with the existing asymmetry argument, meaningfully narrow","revision_made":"no","referee_comment":"§5, Table 3: The argument against a tokenization confound for the cross-dataset transfer asymmetry is made verbally rather than empirically. The paper should characterize tokenization differences across datasets and explain why the asymmetry argument is robust to them, or provide a fixed-tokenizer control."}],"tokens_in":16847,"tokens_out":3080,"duration_ms":65603,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The paper's core finding is real and worth a referee's time: remaining output length is linearly decodable from the residual stream of 7-8B instruction-tuned LLMs, and probes trained on natural-language data transfer cross-dataset in a way that synthetic-trained probes do not. The cross-dataset transfer asymmetry (Tables 3, 9, 10) is the most informative result — it's not something a memorize-the-marginal story easily produces, and the authors are appropriately careful about what decodability does and doesn't imply causally. The three-model, seven-dataset, three-seed design with constant-median and exact-countdown baselines is well-structured. The per-layer sweep (Figure 4) showing the signal emerges in late layers, not at the embedding, is a useful control against trivial token-identity confounds. Code is provided via anonymous repo. The paper is also transparent about its own limitations — unusually so for the third claim, which I'll get to below. Credit is earned on the first two claims and on methodological honesty. The stress-test concern about prompt-content confounds is partially right but overstated. On synthetic datasets (Count/Countdown), the paper already acknowledges the probe is reading n from the prompt — that's the point of including them as a positive control, not as evidence for a plan. The concern is more relevant for natural-language datasets: the constant-median baseline uses zero prompt information, so the probe's improvement could partly reflect content-to-length correlations (harder GSM8K problems produce longer solutions) rather than an internal length estimate. A simple baseline — regressing T on prompt token count or a bag of prompt features — would sharpen the claim. That said, the cross-dataset transfer results partially address this: if the probe were just reading content features, broad transfer across heterogeneous datasets (GSM8K to MMLU-Pro to Countdown) would be harder to explain than a general length-tracking direction. The transfer asymmetry is evidence against the pure content-confound story, though not conclusive. The third claim — dynamic re-estimation at retraction tokens — is the weak point. Five curated examples from the worst-MAE region, where absolute predictions are off by hundreds of tokens, can only support a directional observation. The authors say this plainly in §4.3 and Limitations. It's a teaser, not a result. The aggregate version with length-matched controls is tractable on their existing cache and should either be done or dropped to an appendix footnote. The 'plan-like representation' language in the abstract and conclusion is stronger than what the evidence supports — what they've shown is a decodable length signal with interesting transfer properties, not that the model uses it to plan. They acknowledge this distinction in §6 but the framing elsewhere oversells it. This is a paper for mechanistic interpretability researchers and anyone interested in diagnostic applications (early termination, CoT faithfulness). The first two claims are solid empirical contributions. The third needs more work. I'd send it to review — the core is worth publishing and the gaps are addressable in revision.","headline":"Solid probing study with a real finding; the 'plan' framing outruns the evidence on the third claim, and a content-confound baseline is missing.","tokens_in":17903,"tokens_out":730,"would_cite":true,"duration_ms":103975,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"LLMs linearly encode how many tokens they have left to generate","keywords":[],"falsifier":"Train probes on synthetic and natural-language datasets that share identical tokenizers and BPE boundaries; if the natural-to-synthetic transfer advantage disappears under matched tokenization, the 'plan' interpretation weakens to a tokenizer artifact.","tokens_in":17289,"feed_emoji":"📏","tokens_out":865,"duration_ms":41921,"temperature":0.7,"pith_summary":"This paper asks whether large language models carry an internal estimate of how much of their response remains to be written, or whether response-length consistency is merely a byproduct of token-by-token decoding. Using minimal linear probes trained on frozen hidden states of three 7-8B parameter models across seven datasets, the authors find three converging pieces of evidence that the model's residual stream contains a plan-like representation of remaining output length. First, total response length is linearly decodable from the prompt's final hidden state alone, before any output token is emitted. Second, probe directions trained on natural-language datasets transfer broadly to other datasets including controlled synthetic ones the probe never saw, while the reverse transfer fails -- an asymmetry the authors interpret as natural-language training recovering a general length-tracking direction rather than memorizing a dataset-specific marginal. Third, on curated examples where the model retracts and restarts a partial solution (e.g., emitting 'Wait, let me try again'), the probe's per-position estimate of remaining length jumps upward, a directional update no position-only predictor can reproduce. The authors frame this as approximate estimation of remaining generation length, distinct from exact-counting impossibility results for transformers, and explicitly note that decodability does not prove the model causally uses this signal.","feed_headline":"LLMs linearly encode their remaining output length","feed_subtitle":"Probes recover a plan-like signal from hidden states before generation starts, transferring across datasets and spiking at retractions.","key_machinery":"A linear probe (single linear layer, no nonlinearity) trained on frozen residual-stream hidden states to predict remaining token count at each generation position. Three variants are compared: the Remaining Count Probe (reads hidden states at every position), a constant-median statistical baseline (dataset median at every position), and a Completion Length Probe (trained only on the prompt's last hidden state, then decremented by position). The cross-dataset transfer matrix tests whether the recovered direction is model-general or dataset-specific.","core_discovery":"The residual stream of an instruction-tuned LLM contains a linearly decodable estimate of how many tokens of output remain, readable from the prompt's last hidden state before generation begins, transferring across datasets with different length distributions, and updating directionally upward when the model retracts and restarts a partial solution. The central object is the Remaining Count Probe -- a single linear layer with no nonlinearity that regresses on the residual-stream activation at each completion position to predict the remaining token count. The asymmetry in cross-dataset transfer (natural-language probes generalize to synthetic data but not vice versa) is itself a key finding,,","pith_inferences":[],"forward_implications":["If the length-tracking direction is causal, ablating or steering it at inference time could control how long a model's response runs, enabling budget-aware generation without external heuristics.","A retraction token that does not produce an upward shift in the probe's estimate could serve as a signature of unfaithful chain-of-thought -- the model says 'wait' but does not internally revise its plan.","Prompt-end length estimates exceeding a token budget could trigger early termination before any output is generated, reducing inference cost.","The cross-dataset transfer asymmetry suggests that models trained on more heterogeneous data develop more general internal representations of output structure, which could inform training-data curation for controllable generation.","If the length signal sharpens with scale, frontier models may carry substantially more precise internal plans for their own output than the 7-8B models tested here, with implications for predictability and steerability of long-form generation."],"fun_headline_variants":["LLM hidden states encode how much is left to say","Hidden states reveal how many tokens an LLM has left to generate","Linear probes decode remaining output length from LLM hidden states","LLMs linearly encode remaining generation length","A linear probe reveals LLMs encode their remaining output length"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The 'plan-like' interpretation rests on the cross-dataset transfer asymmetry being genuine and not driven by tokenization differences between natural-language and synthetic datasets; the paper acknowledges that a shared-BPE replication is left to follow-up work and does not empirically rule out this alternative explanation.","fun_headline_variants_meta":{"raw":{"variants":["LLM hidden states encode how much is left to say","Hidden states reveal how many tokens an LLM has left to generate","Linear probes decode remaining output length from LLM hidden states","LLMs linearly encode remaining generation length","A linear probe reveals LLMs encode their remaining output length","LLM hidden states encode remaining response length before generation"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":1318,"prompt_tokens":544,"completion_tokens":774,"prompt_tokens_details":null},"tokens_in":544,"tokens_out":774,"duration_ms":19444,"temperature":1.0,"reasoning_tokens":724,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-07T17:43:46.139339+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Train probes on synthetic and natural-language datasets that share identical tokenizers and BPE boundaries; if the natural-to-synthetic transfer advantage disappears under matched tokenization, the 'plan' interpretation weakens to a tokenizer artifact.","supporting_citations":[],"review_version":1}