{"id":"88d7231a-62f6-4842-b498-7b7141f3e06d","arxiv_id":"2412.07961","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"By re-sampling alternate tokens at each position of generated text, the authors find many 'forking tokens' where a single token choice flips the model's final answer distribution, including chain-of-thought steps and punctuation-like tokens.","lead":"This paper asks whether a single changed token can make a language model produce a completely different answer, and develops a method to find such 'forking tokens' by re-sampling alternate continuations at every position; on GPT-3.5 across seven tasks it finds many abrupt shifts in the model's answer distribution, including at seemingly trivial tokens.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quantitative claim that a majority of sequences have S(T)<0.2 depends on an unvalidated extractor that demonstrably splits one semantic answer into several outcome bins (App. E, Fig. 23), so the headline low survival rates may be partly extractor artifacts.","rationale":"I read the paper in good faith. The Forking Tokens Hypothesis is clearly stated, and the qualitative case studies (HotpotQA-8076, MMLU-12, GSM8k-59) are convincing demonstrations that some token positions carry large outcome consequences. The method is novel and reasonably specified. The load-bearing quantitative step is Sec. 4.2's survival analysis, and the reader correctly identified the extractor's binning as the most fragile link. The paper even discloses the failure mode in App. E but does not quantify its effect; that omission is exactly what makes the concern unresolved. A resolution is cheap: manually merge semantic duplicates and re-run the survival statistics. I do not see an internal inconsistency; the concern is about unrecognized sensitivity of the headline rate to a pipeline component. The qualitative examples and the multiple-choice tasks (where R is clean) provide independent support, so the hypothesis is not simply an artifact; but the 'majority of sequences' claim needs the proposed test. Therefore the reader's CONDITIONAL verdict is the right one and my stress test does not change it.","tokens_in":37364,"tokens_out":7720,"duration_ms":79724,"concrete_test":"Annotate a random sample of R outputs from HotpotQA and LastLetter (e.g., 100 per task) with semantic equivalence labels (same answer under a simple rubric; include aliases, abbreviations, and translations as equivalent). While holding all other pipeline choices fixed (same completions, same epsilon=0.6, same BEAST settings), merge all extracted strings belonging to the same semantic answer class into a single outcome bin, recompute ot,w, d, h(t), S(T), and report the fraction of sequences with S(T)<0.2. If that fraction drops below the reported majority (or below 50% on the annotated subset), the extractor granularity is a primary driver of the headline survival-rate claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative evidence for the Forking Tokens Hypothesis is Sec. 4.2's survival analysis: a majority of the 210 sequences have S(T) < 0.2 at epsilon = 0.6. S(T) is computed from Eq. (3) using indicators d(ot,w, ot,w*) > epsilon, where ot,w are histograms over categorical outcomes R produced by a second LLM (Gemini Flash) with task-specific prompts and Python cleansing (App. G). Every distance computation inherits the extractor's binning. The paper itself shows in App. E (Fig. 23) that for a HotpotQA item the same real-world entity is returned as 'Natalia Lafourcade', 'Nati', 'Nati Lafourcade', and 'Natalia y La Forquetina'; these are distinct strings but not distinct semantic answers. The extractor is never validated (no human agreement, no error analysis), and no code/data are released. If such over-splitting is common, d(ot,w, ot,w*) is inflated, hazards h(t) are overestimated, and S(T) is biased downward, making the 'often just a single token away' conclusion partly an artifact of surface-form variation rather than genuine semantic forks. Because the paper's headline rate is the single most load-bearing quantitative result, this is the key concern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Forking Tokens Hypothesis: that in LLM text generation a single token substitution at certain positions can dramatically change the final outcome. To test this, the authors develop Forking Paths Analysis, a three-stage black-box pipeline: (1) decode a greedy base path and record the top alternate token probabilities; (2) at each token position, resample S=30 completions conditioned on each plausible alternate token; (3) extract a categorical outcome from each completion using a second LLM (Gemini Flash). The resulting per-token outcome distributions ot and ot,w are analyzed with Bayesian change point detection (BEAST) on a univariate drift series, and with a discrete-time survival analysis on per-token hazards. Experiments on 7 tasks across 4 domains with GPT-3.5 report both individual qualitative examples of forking tokens (e.g., HotpotQA-8076 flipping from Robin Tunney to Mia Sara) and aggregate quantitative results: change-point fractions per task and low terminal survival rates S(T), with mean S(T) below 0.2 for a majority of sequences at typical thresholds. The paper concludes that static final-token uncertainty estimates miss substantial dynamical uncertainty and that LLMs are often one token away from a very different answer.","tokens_in":37546,"tokens_out":7726,"duration_ms":80348,"significance":"The qualitative phenomenon documented here is genuinely interesting and, if the quantitative claims are correct, would be an important finding for uncertainty estimation and evaluation of LLMs: it would show that final-answer confidence can be misleading and that outcome uncertainty can collapse or shift abruptly at individual tokens. The method is novel, model-agnostic, black-box, and the interactive dashboard is a useful public artifact. The paper also demonstrates concrete, semantically meaningful forking examples (2021 vs 2024; Robin Tunney vs Mia Sara) that are not explained by simple low-probability sampling noise, and it shows a lack of correlation between token logit probability and change-point probability (App. C.2). However, the paper's load-bearing quantitative claims rest on an unvalidated second-LLM outcome extractor that the authors themselves show splits one semantic answer into several surface forms, and on a survival function that appears mis-specified as written. These issues must be resolved before the extent of the phenomenon can be accepted.","major_comments":[{"comment":"The survival function is defined as S(t) = 1 - ∏_{t'=1}^{t} h(t'), where h(t) is the probability that a sampled alternate token is forking. This is not the standard discrete survival function, which is ∏_{t'=1}^{t} (1 - h(t')). For typical small hazard values, 1 - ∏ h(t') is close to 1 and does not decrease monotonically over a long sequence, yet the plotted S(t) curves in Fig. 7 (Bottom Left) decrease smoothly from 1 to below 0.3, which matches the standard survival definition. As written, the equation is therefore inconsistent with the paper's own plots and with the reported low S(T) values. Please correct Eq. (3) and re-derive the numerical results in Table 1 and Fig. 7 under the intended definition, and state clearly whether S(T) is a survival probability or a cumulative fork probability.","section":"Sec. 4.2, Eq. (3)"},{"comment":"The outcome extractor R maps each continuation to a categorical string, and every distance d(ot,w, ot,w*) in Eq. (3) is computed over these categories. The paper itself documents in App. E (Fig. 23) that the single entity 'Natalia Lafourcade' is returned as 'Natalia Lafourcade', 'Nati', 'Nati Lafourcade', and 'Natalia y La Forquetina' and counted as distinct outcomes. This over-splitting inflates distances, inflates the hazard h(t), and biases S(T) downward. The extractor is never validated against human judgment, and no error analysis is reported. Please provide a human-validation study of R on a random sample of extractions per task, and recompute the survival analysis after collapsing semantic equivalents (e.g., via a secondary aggregation step or a semantic-similarity threshold), reporting both the raw and collapsed estimates.","section":"Sec. 2.4 and App. E (Fig. 23)"},{"comment":"The claim that a majority of sequences have S(T) < 0.2 is reported as a mean over 30 examples per task, with no confidence intervals, no per-task distribution, and no sensitivity analysis for the two most consequential free choices: the distance threshold epsilon (Table 1 uses epsilon = 0.6; Fig. 7 reports 'for all epsilon < 0.9' without error bars) and the alternate-token sampling rule (k <= 10, p >= 5%). Because S(T) is computed on the unvalidated extractor's bins, the quantitative headline may be an artifact of the binning and threshold choices. Please report the full distribution of S(T) per task (e.g., box plots), and include a robustness check with a coarser outcome space (multiple-choice letter only; numeric answers normalized as in the existing cleansing functions) to demonstrate that the conclusion is not driven by the chosen granularity.","section":"Sec. 4.2, Table 1 and Fig. 7"}],"minor_comments":[{"comment":"The sentence 'For (1), we collect N=300 full text responses' refers to an undefined item (1); this is presumably the outcome distribution ot at t=0, but it should be stated explicitly.","section":"Sec. 3, paragraph 2"},{"comment":"The prompt listings contain stray '→' characters and '⁄tildelow' placeholders that appear to be rendering artifacts; these should be cleaned before publication.","section":"App. G"},{"comment":"The text refers to 'logit probabilities' when it appears to mean next-token probabilities after the softmax; please use consistent terminology.","section":"Sec. 2.1"},{"comment":"The semantic drift transformation yt = d(o0, ot) uses L2 distance on probability histograms; the choice of distance metric is a free parameter and the justification for L2 over other metrics (e.g., total variation or Jensen-Shannon) should be expanded.","section":"Sec. 2.3"},{"comment":"The reproducibility statement says code and data 'will be made publicly available in the near future'; for a methods paper, a stable repository link or a specific commitment is needed.","section":"Reproducibility statement"}],"recommendation":"major_revision","confidential_remarks":"The paper has a strong qualitative core: the concrete forking examples are striking and the method is original. The main risk is that the quantitative survival-rate claims are not yet supported because the outcome extractor is unvalidated and the survival function appears mis-specified in Eq. (3). I would like to see the authors address those two points head-on, ideally with a human evaluation of R and a corrected/recomputed survival analysis. If the extractor artifact is severe, the 'majority of sequences have S(T)<0.2' claim may weaken substantially, but the qualitative existence of forking tokens would likely survive. This fits the journal's scope well once the validation gap is closed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is good: the per-token counterfactual re-sampling pipeline and the outcome-distribution time series are genuinely new relative to static uncertainty estimation work like Kadavath, Tian, and Xiong. The concrete examples are compelling—Robin Tunney versus Mia Sara, 2021 versus 2024, the parenthesis token in GSM8k—and they make a real case that forking tokens exist. The paper is also clearly written and admirably honest about its own limitations, including the extractor issue in App. E.\n\nThe soft spot is the quantitative extent claim. Survival rates in Sec. 4.2 come from Eq. (3), where a fork is declared when d(ot,w, ot,w*) > epsilon, and that distance is computed over histograms of extracted outcomes. The extractor is a second LLM with Python cleansing, never validated against human judgment. App. E, Fig. 23 shows the same person returned as 'Natalia Lafourcade', 'Nati', 'Nati Lafourcade', and 'Natalia y La Forquetina'—four bins for one answer. If that over-splitting is at all common, L2 distances inflate, hazards rise, and S(T) is biased downward. So the 'majority of sequences have S(T) < 0.2' headline is not yet supported. The post hoc tuning of the change-point detector (noise injection, alpha, epsilon chosen to match qualitative judgment) and the absence of released code or data make those numbers even harder to check.\n\nThat said, the existence of forking tokens does not rest on the extractor. The case studies show real semantic flips at individual tokens, and even if the true prevalence is lower than reported, the phenomenon is real and the method is a useful tool. The correlation analysis in App. C.2 (token logits do not predict change points) is a nice sanity check.\n\nThis paper deserves a serious referee. The referee should press for extractor validation or a similarity-based distance that does not treat surface variants as distinct outcomes, for sensitivity analysis on epsilon and binning, and for released code and data. As it stands, the qualitative contribution is solid, but the headline quantitative claims should be treated with caution.","headline":"A genuinely new method and a plausible hypothesis, but the headline prevalence numbers are undercut by an unvalidated extractor that demonstrably over-splits surface variants; the qualitative existence claim holds.","tokens_in":38201,"tokens_out":2326,"would_cite":true,"duration_ms":24847,"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":"A single token swapped for a probable alternative can redirect an LLM's entire subsequent generation, and static uncertainty estimates miss this.","keywords":["forking tokens","uncertainty estimation","large language models","change point detection","survival analysis","text generation"],"falsifier":"Take the HotpotQA example in Figure 23 and re-run Forking Paths Analysis with an extractor that merges surface variants of the same person ('Natalia Lafourcade', 'Nati', 'Natalia y La Forquetina') into one semantic outcome. If the change point probability and survival hazards at the separating tokens drop below significance, the apparent fork is an extraction artifact. A broader check: if merging such variants across all HotpotQA sequences moves the majority of end-of-sequence survival rates above 0.2, the headline claim of widespread low survival would not survive its own measurement assumptions.","tokens_in":37027,"feed_emoji":"🔀","tokens_out":7135,"duration_ms":66755,"temperature":0.7,"pith_summary":"This paper sets out to show that uncertainty in large language models is not just a property of the final answer: it lives at individual tokens along the way. The authors propose the Forking Tokens Hypothesis, namely that replacing one token with a probable alternative can change the distribution of everything that follows. They test it by re-sampling many completions from every likely alternate token in GPT-3.5 outputs across seven tasks, extracting final answers with a second LLM, and using change point detection and survival analysis to find where outcome distributions break. They report forking tokens in content words, spaces, and punctuation, and find that a majority of analyzed sequences have a low probability of surviving to the end without a major outcome shift. If correct, the result means existing token-level or final-token confidence scores can be systematically misleading, and safety evaluations that sample one response per prompt can miss large parts of the model's actual behavior.","feed_headline":"One token swap can flip an LLM's answer","feed_subtitle":"A token-level analysis finds LLMs often one probable token away from a very different answer.","key_machinery":"The machinery is Forking Paths Analysis, built on the Forking Tokens Hypothesis. For a greedily decoded base path $x^*$, the pipeline records the top probable alternate tokens at each index $t$, re-samples fresh completions conditioned on each alternate token, and maps every completion to a categorical outcome $R$ using a second LLM. These outcomes are aggregated into probability-weighted histograms $o_{t,w}$, and the marginal distribution $o_t$ becomes a multivariate time series. A one-dimensional 'semantic drift' series $y_t = d(o_0, o_t)$ with an $L^2$ distance feeds a Bayesian change point detection model that flags indexes where the outcome distribution shifts abruptly; a separate discrete-time survival analysis computes, for each token, a hazard $h(t) = \\sum_w p(x_t=w \\mid x_{<t}) \\mathbb{1}[d(o_{t,w}, o_{t,w^*})>\\epsilon]$ and the cumulative survival $S(t)$, so a 'forking token' is an index or alternate value where a small probable substitution changes the downstream answer distribution.","core_discovery":"The paper's central claim is empirical support for the Forking Tokens Hypothesis: in LLM text generation there exist token positions where exchanging the sampled token for a probable alternate significantly changes the distribution of eventual outcomes. The authors demonstrate this by building per-token outcome distributions from re-sampled completions and analyzing them with Bayesian change point detection and survival analysis. They find forking tokens across all seven tasks, including seemingly innocuous tokens such as spaces, parentheses, and the word 'that' versus 'who'. In many sequences the outcome distribution is stable for most of the text and then collapses to a different answer at a single forking token, often a hallucinated reasoning step. Their survival analysis shows that for a large fraction of GPT-3.5 sequences the cumulative probability of reaching the end of the base path without hitting a fork is below 0.2, even with a generous distance threshold, so the paper concludes that static final-token confidence estimates can be unstable and misleading.","pith_inferences":["If forking tokens generalize across models, the per-token hazard profile could be used as a model diagnostic: where in a prompt a model commits to an answer, and how sharply.","A natural extension is to use semantic embeddings instead of categorical extraction, which would let the same machinery measure forking in open-ended creative text rather than only tasks with extractable answers.","The survival rate $S(T)$ could be turned into a calibration-style metric: of two models with equal accuracy, the one with higher survival rate is more stable under token-level perturbations.","Forking positions may be a target for adversarial token edits: an attacker who can identify probable alternate tokens at high-hazard positions can steer the model's outcome with minimal surface changes."],"forward_implications":["Static final-token confidence, such as a near-100% logit at the last token, can hide a sequence that was close to a different answer for most of its length.","Evaluators should treat low survival rates as a warning that a single sampled completion is not a stable measurement of model knowledge.","Safety and alignment checks that sample one or a few responses can miss paths reachable by probable token substitutions, including hallucinated or harmful completions.","Token logit probability is not a reliable flag for forking: punctuation and space tokens can carry as much outcome-changing weight as content words.","The analysis works without fine-tuning or model weights, so it can be applied to black-box APIs and to any task where final answers or semantic embeddings can be extracted."],"supporting_citations":[{"why":"Supplies the Bayesian change point detection implementation used to locate abrupt shifts in outcome distributions.","marker":"Zhao, 2019"},{"why":"Provides the zero-shot chain-of-thought prompt and answer cleansing functions used across tasks.","marker":"Kojima et al., 2022"},{"why":"Contributes the GSM8k math word-problem dataset used to measure forking in open-ended mathematical answers.","marker":"Cobbe et al., 2021"},{"why":"Contributes the MMLU multiple-choice benchmark used for forking analysis.","marker":"Hendrycks et al., 2020"},{"why":"Provides HotpotQA multi-hop questions, including the paper's main worked example of a hallucination fork.","marker":"Yang et al., 2018"},{"why":"Supplies StoryCloze, the open-ended story generation task with two-choice endings.","marker":"Mostafazadeh et al., 2017"},{"why":"Represents the prior static uncertainty approach that the paper argues misses token-level forks.","marker":"Kadavath et al., 2022"}],"fun_headline_variants":["One token can fork an LLM's entire answer","LLMs often one token away from a different answer","Forking tokens: tiny swaps, big outcome shifts","Punctuation can change an LLM's whole story"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results assume the second model used to extract final answers treats surface variants as the same answer whenever they are semantically the same answer; if it does not, many detected 'forks' and low survival rates could be artifacts of word choice rather than real changes in meaning.","fun_headline_variants_meta":{"raw":{"variants":["One token can fork an LLM's entire answer","LLMs often one token away from a different answer","Forking tokens: tiny swaps, big outcome shifts","Punctuation can change an LLM's whole story"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00018,"raw_usage":{"total_tokens":1281,"prompt_tokens":902,"completion_tokens":379,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":314}},"tokens_in":518,"tokens_out":379,"duration_ms":3648,"temperature":1.0,"reasoning_tokens":314,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:22:14.378935+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the HotpotQA example in Figure 23 and re-run Forking Paths Analysis with an extractor that merges surface variants of the same person ('Natalia Lafourcade', 'Nati', 'Natalia y La Forquetina') into one semantic outcome. If the change point probability and survival hazards at the separating tokens drop below significance, the apparent fork is an extraction artifact. A broader check: if merging such variants across all HotpotQA sequences moves the majority of end-of-sequence survival rates above 0.2, the headline claim of widespread low survival would not survive its own measurement assumptions.","supporting_citations":[],"review_version":1}