{"id":"d387f65b-cb83-4d9f-907b-0d48cb763f39","arxiv_id":"2412.11937","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Adding a reversed, scaled sinusoidal positional encoding to a fine-tuned decoder-only LLM lets it end responses within about three tokens of a requested length.","lead":"This paper adds a countdown-style positional encoding to decoder-only language models, then fine-tunes them so they stop generating at a user-chosen token length. If it holds up, it would give production systems exact control over response length without changing the model architecture.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No control condition isolates the reversed countdown from any added per-token signal, so the claim that LDPE itself enables length control is not independently established.","rationale":"The reader's weakest assumption already identifies that the countdown's causal role is not independently established; my stress-test sharpens this into a missing control condition. The paper's experiments compare LDPE against prompt-based length control, but that baseline changes multiple factors at once, so it cannot rule out the possibility that any salient added per-token signal would produce similar length fidelity after fine-tuning. This is load-bearing because the central claim is not merely that the method works, but that the reversed countdown is the mechanism. The proposed concrete test would settle this by holding the fine-tuning pipeline fixed while varying only the semantics of the added signal. If forward or random encodings also achieve low error, the method's novelty and explanation weaken; if LDPE uniquely succeeds, the mechanism is supported. The evaluation-size and selection-bias concerns raised by the reader also remain, but they are secondary to this causal-control gap. Since the paper still reports plausible results and the concern can be addressed with additional experiments, the CONDITIONAL verdict remains appropriate without escalation to rejection.","tokens_in":11527,"tokens_out":7473,"duration_ms":76405,"concrete_test":"Fine-tune Mistral-7B-Instruct on the same 110k samples with identical LoRA settings under four conditions: (1) LDPE as in the paper; (2) forward (non-reversed) sinusoidal positional encodings; (3) random fixed per-position vectors; (4) no added encoding, with target length given in the prompt. Choose hyperparameters on a separate validation split, then evaluate on a held-out set of 1,000 OpenOrca/MMLU prompts at target lengths 10-200, reporting mean absolute token error and 95% confidence intervals per length bin. If condition (2) or (3) also yields mean error below 3 tokens, the reversed countdown is not the operative mechanism; if condition (1) uniquely achieves low error, the central LDPE claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the reversed sinusoidal index acts as a portable remaining-token-budget signal. The paper's experiments never include a control that keeps the added-embedding manipulation identical while changing the countdown semantics. The only length-control baseline is prompt-based fine-tuning (Section 4), which differs in multiple ways: no added per-token signal, and the target length is given as text. Consequently, the observed low mean token error could be produced by a length prior in the 110,000-sample training set, by the model exploiting any consistent positional feature, or by the specific reversed countdown. The 20 hand-selected QA prompts and GPT-3.5-summary comparisons do not separate these hypotheses, and hyperparameters were tuned until satisfactory results (Section 4), so selection bias is a live threat. Without a control such as forward sinusoidal or random fixed per-position encodings under identical fine-tuning, the causal role of LDPE is not independently established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for precise token-level length control in decoder-only LLMs. It adapts a reverse sinusoidal positional encoding, called LDPE (and an offset variant ORPE), adds it to the input token embeddings, and fine-tunes the model with LoRA. At inference, the encoding counts down from a user-specified target length, and the model is expected to terminate at the set length. The authors report mean absolute length errors below three tokens on a summarization task, qualitative evidence on 20 question-answering prompts, a Max New Tokens++ extension for upper-bound control, and benchmark results indicating preserved general quality. The paper is clearly written and the method is simple, but the empirical evidence is thin in several respects, and a missing control leaves the causal role of the countdown direction unestablished.","tokens_in":11685,"tokens_out":4724,"duration_ms":42762,"significance":"If the central claim is correct, the method provides a lightweight, LoRA-based way to give decoder-only LLMs explicit token-budget awareness, which is practically valuable for production systems requiring structured or length-constrained outputs. The approach builds on existing encoder-decoder length-control ideas (Takase and Okazaki, 2019) and adapts them to modern decoder-only models, and the Max New Tokens++ extension is a plausible way to relax exact-length targets. The paper uses standard open models (Mistral-7B, Llama3-8B) and standard datasets (OpenOrca, MMLU, CNN/DailyMail), which is a strength for reproducibility, and the limitations section is candid about several remaining gaps. However, the evidence base is too narrow to fully support the claims: the control condition problem, the tuning-on-evaluation issue, and the small hand-selected QA set all bear directly on the paper's central assertions.","major_comments":[{"comment":"The claim that the reversed countdown semantics of LDPE/ORPE are what enables length control is not isolated by the experiments. The only length-control baseline is prompt-based fine-tuning, which differs from LDPE in multiple ways: it adds no per-token encoding, the target length is textual, and the model receives the target only once. The observed low errors could therefore be due to the added per-token signal generally (e.g., any consistent positional feature), to a length prior learned from the 110,000-sample training set, or to the countdown specifically. A control condition using the same added-embedding procedure but with forward sinusoidal encodings, or with fixed random per-token encodings, under identical fine-tuning is necessary to attribute the effect to the countdown direction. Without such a control, the causal role of LDPE is not independently established.","section":"Section 3.1 and Section 4"},{"comment":"The manuscript states that hyperparameters were \"obtained via manual experimentation and evaluation until satisfactory results were obtained.\" No separate held-out model-selection split is described; the only evaluation set consists of 200 samples drawn in the same manner as the training set, and the QA length-control results use 20 hand-selected prompts. This means the reported numerical results are likely optimistically biased, because the same evaluation data may have been used to guide tuning. The paper should report results on a genuinely held-out test split, describe the tuning procedure, and provide error bars (e.g., bootstrap confidence intervals or multiple training seeds) for the mean token errors.","section":"Section 4, Hyperparameters"},{"comment":"The question-answering length-control evaluation is based on only 20 hand-selected samples, and the results are presented solely as scatter plots. No quantitative summary (mean or median absolute error, standard deviation, or per-target-length breakdown) is given, and the selection criterion for the 20 prompts is not stated. The visual claim of \"near-perfect alignment\" is not sufficient to support the abstract's \"mean token errors of less than 3 tokens,\" which is actually only supported by the summarization experiment. The authors should provide numeric length-error statistics for both models and for ORPE, along with confidence intervals and a description of how the 20 prompts were chosen.","section":"Section 5.1.1"},{"comment":"The response-quality claim \"without compromising response quality\" is primarily evaluated by comparing generated summaries to GPT-3.5-turbo-generated summaries using BERT and ROUGE scores. This is an automatic proxy, not a human judgment, and the target summaries themselves are generated by another LLM, so the comparison measures similarity to a particular model's style and content rather than absolute quality. The standard benchmarks in Section 5.1.3 partially mitigate this, but one benchmark (Mistral on HellaSwag) shows a reduction, and the evaluation harness was modified for QA tasks. A human evaluation of summary faithfulness, or at least evaluation against a dataset with human-written reference summaries, would substantially strengthen the quality-preservation claim.","section":"Section 5.1.2"},{"comment":"The description of the LM Evaluation Harness modification is insufficiently precise. The text says the harness was \"modified to optionally apply the LDPE countdown to the queries\" for QA tasks, and that this gave the model \"additional information about the response length, but no information about whether a specific question response pair was correct or incorrect.\" It is unclear whether the base (non-fine-tuned) models also received these added LDPE encodings during evaluation. If they did, adding untrained encodings could distort their log-likelihoods; if they did not, the comparison is not controlled. The exact evaluation protocol for each model and task should be specified in detail.","section":"Section 5.1.3"}],"minor_comments":[{"comment":"The abstract's claim of \"mean token errors of less than 3 tokens\" is presented without specifying that it comes from the summarization experiment; the QA experiment does not report a mean error. The scope of this claim should be stated explicitly.","section":"Abstract and Section 5.1.1"},{"comment":"The Max New Tokens++ results are shown only as a plot with no numeric metrics. The statement that the 95% confidence interval \"still follows closely\" should be defined (e.g., bootstrap over samples) and quantified with a table of, for example, median absolute error and coverage of the confidence interval.","section":"Section 5.2"},{"comment":"The table caption says \"ORPE fine tuned Llama3 8B\" while Section 5.1.1 reports the LDPE results in the main text; please verify that the model and encoding names are consistent throughout.","section":"Appendix A, Table 1"},{"comment":"The OpenOrca reference contains a malformed URL (\"https://https://huggingface.co/...\"). Please correct the citation.","section":"Section 4, Data"},{"comment":"There are minor grammatical issues, for example \"for the Mistral-prompted was an order of magnitude larger\" and \"the target length as was added to the prompt.\" A copy-editing pass is recommended.","section":"Section 5.1.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is simple and the direction is interesting, but the empirical evaluation is not yet at the level that would justify the strong claims in the abstract. The two issues I consider load-bearing are (i) the absence of a control that isolates the countdown semantics from the mere addition of per-token encodings, and (ii) the apparent tuning on the evaluation data. Both are addressable within the scope of a revision. The paper is a reasonable fit for a CS.CL venue, though the novelty relative to prior length-control work is incremental."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper shows that adding a reversed sinusoidal positional encoding to input embeddings and LoRA fine-tuning a 7–8B instruct model gives token-level length control with mean errors around 2–3 tokens. That is a real, useful result. But the evaluation is thinner than the headline, and the paper’s countdown interpretation is not tested against controls.\n\nWhat’s new: the adaptation of Takase and Okazaki’s LDPE from encoder-decoder to decoder-only LLMs, the ORPE offset variant, the Frobenius-norm scaling trick, and a preliminary Max New Tokens++ extension. None of these are paradigm-shifting, but they are sensible engineering contributions. The paper also does a few things well: it evaluates quality on standard benchmarks plus BERT/ROUGE against GPT-3.5 summaries, and the quality numbers are essentially flat across models, which supports the claim that LDPE doesn’t degrade responses. The limitations section is refreshingly candid, including that ORPE gives no improvement and that Max New Tokens++ is preliminary.\n\nNow the soft spots, in proportion. The central empirical claim is probably correct in its narrow form – the method does control length. But the evaluation used 20 hand-selected QA prompts with no error bars, hyperparameters were tuned until the results looked good, and no code or data are released. That alone would make me want more evidence before trusting the precise <3-token error. The deeper issue is the stress-test concern: there is no control that keeps the added-embedding manipulation identical while changing the countdown semantics. Comparing LDPE against prompt-based fine-tuning differs in two ways at once – the added per-token signal and the target-length-as-text. So the paper has not shown that the reversed countdown, rather than any consistent per-token positional feature, is what drives the length control. The claim that the model “learns a token budget” is an interpretation, not a demonstrated mechanism. That said, this is fixable: add controls with forward sinusoidal or random fixed per-position encodings under identical fine-tuning. This concern weakens the mechanism story but does not sink the method.\n\nWho is this for? Anyone working on length control, decoding strategies, or positional encodings for LLMs will find it worth a read. It is not a landmark paper, but it is a competent engineering result with honest reporting. I would send it to peer review – a serious referee can push for the control experiment, a larger evaluation, and artifact release. The paper deserves that engagement rather than a desk reject.","headline":"Solid, honest adaptation of LDPE to decoder-only LLMs with clear length-control gains, but the evaluation is thin and the countdown mechanism is not isolated from alternative positional signals.","tokens_in":12223,"tokens_out":2110,"would_cite":true,"duration_ms":21381,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Decoder-only LLMs can be given precise, token-level response-length control by adding a reversed positional encoding that counts down to the target end; low-rank fine-tuning yields mean token errors under 3 tokens.","keywords":["length control","positional encoding","decoder-only language models","token budget","low-rank adaptation","summarization","question answering","Max New Tokens++"],"falsifier":"Evaluate an LDPE fine-tuned model on a large held-out prompt set at target lengths of 5, 300, and 1000 tokens, and compare the mean absolute token error against lengths inside the 10-200 training range; if the error grows steeply outside the range, the model is memorizing training-length associations rather than using the countdown as a general remaining-token budget.","tokens_in":11310,"feed_emoji":"🎯","tokens_out":5770,"duration_ms":50477,"temperature":0.7,"pith_summary":"The paper claims that decoder-only large language models can be given precise, token-level control over response length without hurting quality. The mechanism is a second positional encoding, added to the token embeddings, that counts down from the desired response length; after a one-epoch low-rank fine-tune, the model learns to end its response when the countdown reaches zero. On question answering and summarization, the authors report mean token errors below 3 tokens, with summary quality scores comparable to a prompt-fine-tuned baseline. The practical value is that production systems needing answers of a specified length could use fine-tuning rather than prompt engineering or hard truncation.","feed_headline":"Countdown encoding steers LLM length to within 3 tokens","feed_subtitle":"Adding a reversed positional encoding during fine-tuning lets decoder-only models end responses at a user-set length.","key_machinery":"The length-difference positional encoding (LDPE): a sinusoidal positional encoding, $PE(i,2k)=\\sin(i/10000^{2k/d})$ and $PE(i,2k+1)=\\cos(i/10000^{2k/d})$, added to the token embeddings with the index reversed so that $i_{\\mathrm{LDPE}}=(L+1)-i$ gives each token a countdown of remaining tokens to the end. The added signal is rescaled by the ratio of the Frobenius norms of the token embeddings and the positional encodings, so it does not drown out the learned embeddings, and a low-rank fine-tune teaches the model to terminate at the countdown's end. The offset variant (ORPE) applies the countdown only to the response portion; Max New Tokens++ adds a random half-normal shift to the countdown during training so the model can stop before an upper bound.","core_discovery":"Adding a second, reversed sinusoidal positional encoding to input token embeddings, with an optional offset so the countdown begins only at the response, teaches a decoder-only LLM during low-rank fine-tuning to associate the remaining encoding index with the number of tokens left to generate. At inference, setting the encoding to a desired total length yields responses that terminate near that length; the paper reports mean absolute token errors of 2.8 and 2.4 tokens on summarization for the two tested 7-8B parameter instruction-tuned models, and near-perfect length alignment on question-answering targets from 10 to 200 tokens, while BERT and ROUGE scores match a prompt-fine-tuned baseline.","pith_inferences":["If the remaining-budget interpretation is correct, the same countdown should transfer to word- or character-level targets by rescaling the index; the paper itself lists this as future work.","The conditioning signal is not inherently tied to the end-of-sequence token, so the same mechanism might steer other discrete choices, such as how many facts to state or where to place a citation.","A stronger test would evaluate out-of-range lengths, such as 5, 300, or 1000 tokens, on a large held-out prompt set; the paper's current evidence does not fully isolate the countdown from memorized length associations.","Replacing the sinusoidal countdown with a learned scalar 'remaining budget' embedding could preserve control while avoiding interference with existing positional encodings, and would be a direct way to test whether the countdown or the fine-tune itself carries the effect."],"forward_implications":["Question-answering responses can be generated at any user-chosen length between 10 and 200 tokens with near-perfect target alignment.","Summarization reaches a mean absolute length error of 2.4-2.8 tokens, versus roughly 24.8 tokens for a prompt-based length-control baseline, at comparable BERT and ROUGE quality.","The method works on existing decoder-only models without changing the pretrained rotary positional embeddings, so it can be added as a low-rank fine-tune rather than a full retraining.","Max New Tokens++ extends exact-length control to upper-bound control, allowing the model to end before the limit when its response is naturally complete."],"supporting_citations":[{"why":"Supplies the length-difference positional encoding that the paper adapts from encoder-decoder models to decoder-only models.","marker":"[27]"},{"why":"Provides the low-rank adaptation method used to fine-tune without overwriting pretrained weights.","marker":"[10]"},{"why":"One of the two decoder-only instruction-tuned models on which the method is demonstrated.","marker":"[11]"},{"why":"The other base model tested for exact and upper-bound length control.","marker":"[20]"},{"why":"Supplies 100,000 question-answer fine-tuning samples used to teach the countdown.","marker":"[15]"},{"why":"Adds 10,000 multi-task samples to the fine-tuning set for topic and length diversity.","marker":"[9]"},{"why":"Provides the news articles whose machine-generated summaries define the summarization targets.","marker":"[21]"}],"fun_headline_variants":["LLMs learn to stop on cue with countdown encoding","Stop on a dime: LLM length control via LDPE","Precise token budgets: LLMs count down to finish","Countdown encoding: precise length control for LLMs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reverse positional encoding is learned as a general 'remaining token budget' that transfers from the 110,000-sample fine-tuning set to unseen prompts and to target lengths across the 10-200 token range, even though the evaluation used 20 hand-selected question-answering prompts and no held-out model-selection split.","fun_headline_variants_meta":{"raw":{"variants":["LLMs learn to stop on cue with countdown encoding","Stop on a dime: LLM length control via LDPE","Precise token budgets: LLMs count down to finish","Countdown encoding: precise length control for LLMs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000744,"raw_usage":{"total_tokens":3264,"prompt_tokens":839,"completion_tokens":2425,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":2357}},"tokens_in":455,"tokens_out":2425,"duration_ms":15751,"temperature":1.0,"reasoning_tokens":2357,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:26:05.762586+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate an LDPE fine-tuned model on a large held-out prompt set at target lengths of 5, 300, and 1000 tokens, and compare the mean absolute token error against lengths inside the 10-200 training range; if the error grows steeply outside the range, the model is memorizing training-length associations rather than using the countdown as a general remaining-token budget.","supporting_citations":[{"cited_title":"Positional encoding to control output se- quence length","cited_arxiv_id":null,"evidence_quote":"Supplies the length-difference positional encoding that the paper adapts from encoder-decoder models to decoder-only models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"One of the two decoder-only instruction-tuned models on which the method is demonstrated."},{"cited_title":"Introducing Meta Llama 3: The most capable openly available LLM to date — ai.meta.com","cited_arxiv_id":null,"evidence_quote":"The other base model tested for exact and upper-bound length control."},{"cited_title":"Openorca: An open dataset of gpt augmented flan reasoning traces","cited_arxiv_id":null,"evidence_quote":"Supplies 100,000 question-answer fine-tuning samples used to teach the countdown."},{"cited_title":"Abstractive text summarization using sequence-to- sequence rnns and beyond, 2016","cited_arxiv_id":null,"evidence_quote":"Provides the news articles whose machine-generated summaries define the summarization targets."}],"review_version":1}