{"id":"11602038-ea37-4376-a5af-129f278b93ff","arxiv_id":"2411.10153","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"BONE is a unifying framework that expresses many existing Bayesian online learning methods as combinations of five design choices, plus a new runlength-based method that handles both gradual and sudden changes.","lead":"This paper proposes BONE, a framework that unifies many Bayesian methods for online learning in changing environments by breaking them into five modeling and algorithmic choices. It also introduces a new method within that structure and shows it often predicts better than existing baselines in forecasting, classification, bandit, and segmentation experiments.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RL[1]-OUPR* uses the current observation twice: Eqs. (19)-(20) build the prior from ν_t = p(r_t|D_{1:t}), then Algorithm 4 applies the same y_t as a likelihood, violating the BONE definition of M.3.","rationale":"The reader's verdict is CONDITIONAL and identifies the new method's ε threshold as the weakest assumption. My stress-test agrees that the new method is the fragile part of the central claim, but I find a more fundamental problem: the prior in Eqs. (19)-(20) is built from ν_t = p(r_t | D_{1:t}), which includes the current observation y_t, and then the same y_t is used as the likelihood in Algorithm 4 line 16. This violates Table 1's definition of the M.3 prior as a function of D_{1:t-1} only, so RL[1]-OUPR* is not formally an instance of BONE as defined. This is a correctness risk for the paper's third contribution ('it allows us to propose new methods') rather than a merely stylistic weakness. The unifying-framework contribution itself is not invalidated: the five-choice structure can still organize existing methods, and the issue is local to the new algorithm. A single ablation replacing ν_t with the predictive runlength probability before the observation would settle whether the reported experimental gains are due to double-counting. If the double-counting is material, the paper needs either a corrected version of RL[1]-OUPR* or an explicit statement that this is an ad-hoc non-Bayesian variant outside the formal BONE definition. That is an addressable revision, so conditional acceptance remains the appropriate verdict, with the clarification/ablation as a required condition.","tokens_in":33837,"tokens_out":10739,"duration_ms":108358,"concrete_test":"Re-run experiments in Sections 4.1.1-4.1.3 and 4.2 with one modification: replace the mixing weight ν_t(r^(1)) in Eqs. (19)-(20) and Algorithm 4 lines 8-9 by the predictive runlength probability p(r^(1)_t | D_{1:t-1}) = (1-κ), while keeping the reset decision based on ν_t. If RL[1]-OUPR*'s reported gains over RL[1]-PR and CPP-OU shrink or reverse, the published results are driven by double-counting y_t rather than by the proposed mechanism.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central new-method claim rests on RL[1]-OUPR* being a coherent instance of BONE, but its defining update is internally inconsistent with the framework. Table 1 defines the M.3 prior as π_t(θ_t; ψ_t) := π(θ_t; ψ_t, D_{1:t-1}), i.e., it may depend only on data before the current observation. However, Eqs. (19)-(20) set the prior mean and covariance using ν_t(r_t) = p(r_t | D_{1:t}), which is computed from the current observation y_t via the predictive likelihoods in Algorithm 4 lines 2-4. Algorithm 4 then uses the same y_t again in line 16 to update θ_t from this prior. Thus y_t enters twice: once to decide the runlength/mixing weight and once as the likelihood. This is not a standard two-stage Bayes update, and it makes RL[1]-OUPR* not an instance of BONE as formally defined. If the experimental advantage of RL[1]-OUPR* over one-pass baselines largely disappears when the double-counting is removed, the paper's claim that BONE generates viable new methods is weakened; if the advantage persists, the issue is less severe but the definition still needs revision. The reader's concern about the heuristic ε threshold is related but secondary: even with ε correctly tuned, the update is not a BONE instance unless the prior dependence on y_t is removed or explicitly defended as a non-Bayesian algorithmic choice.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BONE, a five-component framework for online probabilistic learning in non-stationary environments, with modelling choices (M.1 measurement model, M.2 auxiliary process, M.3 conditional prior) and algorithmic choices (A.1 posterior approximation, A.2 weighting for the auxiliary variable). It shows how many existing methods from filtering, changepoint detection, continual learning, and bandits can be expressed as BONE instances (Table 3), and introduces a new method, RL[1]-OUPR*, which combines a single runlength hypothesis with an Ornstein-Uhlenbeck drift and a threshold-based reset. The framework is evaluated on electricity load forecasting, synthetic classification with drift and jumps, Bernoulli bandits, piecewise autoregression, and heavy-tailed regression.","tokens_in":34121,"tokens_out":7848,"duration_ms":76030,"significance":"The modular taxonomy is a genuinely useful organizing contribution: Table 3 convincingly maps a broad literature into a small number of components, and the accompanying JAX library is a concrete, reproducible artifact. The recursions in Section 2 are correct, and the generalization from Bayesian posteriors to loss-based updates and ad-hoc weights is coherent and clearly explained. If the proposed method were a clean instance of the framework and the empirical claims were supported by repeated trials, the paper would be a valuable reference for both practitioners and researchers. The main reservations concern the internal consistency of RL[1]-OUPR* with the formal BONE definition, and the strength of the stochastic classification evidence.","major_comments":[{"comment":"The M.3 prior is formally defined as π_t(θ_t; ψ_t) := π(θ_t; ψ_t, D_{1:t-1}) in Table 1, i.e., it may depend only on data before the current observation. However, Eqs. (19)-(20) and Algorithm 4 lines 2-16 construct the prior using ν_t(r_t) = p(r_t | D_{1:t}), which is computed from the current observation y_t through the predictive likelihoods in Algorithm 4 lines 2-4. Algorithm 4 then multiplies this prior by the same y_t as a likelihood in line 16. Thus y_t is used twice: once to set the runlength/mixing weight and once as the observation in the parameter update. This makes RL[1]-OUPR* inconsistent with the BONE definition in Table 1 and with the generalized posterior update in Eq. (7). Please either revise the definition of M.3 (and the update in Eq. (7)) to permit priors that depend on D_{1:t}, or present RL[1]-OUPR* as a heuristic extension of BONE rather than an instance of it; if the latter, the abstract's claim that BONE 'allows us to propose new methods' needs qualification.","section":"Section 2.5, Eqs. (19)-(20); Algorithm 4; Table 1"},{"comment":"The classification experiments use a stochastic data-generating process (Bernoulli observations in both cases, and random parameter jumps in Section 4.1.3), yet the results are reported as single-run misclassification rates without repeated trials or error bars. The statements that RL[1]-OUPR* 'works the best' and 'significantly outperforms' RL[1]-PR are therefore not statistically supported. Please add multiple seeds with means and confidence intervals, and report how hyperparameters, including the threshold ε, were selected for each method.","section":"Sections 4.1.2 and 4.1.3, Figures 7 and 9"},{"comment":"The threshold ε is a new free parameter of RL[1]-OUPR*, but the paper provides no principled rule for setting it beyond the limits ε=1 (always reset, no learning) and ε=0 (OU-type update). Since a misspecified ε collapses the method to either never adapting or always resetting, the experiments should state the ε values used and include a sensitivity analysis. This is necessary to assess whether the reported advantages of RL[1]-OUPR* over one-pass baselines are robust or are driven by a favorable threshold choice.","section":"Section 2.5, Eqs. (19)-(20); Sections 4.1-4.3"}],"minor_comments":[{"comment":"The notation µ(r_t) and Σ(r_t) in the ν_t(r_t) > ε branch is inconsistent with Eq. (18) and with Algorithm 4 line 8, where the previous posterior (µ_{t-1}, Σ_{t-1}) is used; please align the notation.","section":"Eqs. (19)-(20)"},{"comment":"The heading 'Details of BONE' is followed by no text; either fill in the section or remove the heading.","section":"Section 2.2"},{"comment":"The caption says 'rolling relative absolute error' while the text in Section 4.1.1 says 'rolling mean absolute error (MAE)'; these should be made consistent.","section":"Figure 3 caption"},{"comment":"The phrase 'between March 4 2020 and March March 8 2020' contains a duplicated 'March'; please fix the typo.","section":"Section 4.1.1"},{"comment":"The cardinality for CPT appears as '2 t' and is presumably 2^{t+1}; the value notation '2{0,1,...,t}' also needs an explicit explanation.","section":"Table 2"},{"comment":"Given the preceding 'if ν(r_t^{(1)}) > ε' branch, the condition 'else if ν(r_t^{(1)}) ≤ ε' is equivalent to 'else'; this is not an error, but the asymmetry may confuse readers.","section":"Algorithm 4, line 10"}],"recommendation":"major_revision","confidential_remarks":"The double-counting issue is the main correctness risk and is load-bearing for the paper's claim that BONE generates viable new methods. It is fixable by reframing RL[1]-OUPR* as a heuristic extension or by generalizing the definition of M.3, and the taxonomy itself is valuable enough to warrant a major revision rather than rejection. I would also ask the authors to verify that the hyperparameter comparisons in Sections 4.1.2 and 4.1.3 are fair, since the new method's threshold ε is an additional tuning knob not present in the baselines."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know: first, BONE is a genuinely useful organizing framework, and the paper's mapping of a scattered literature onto five choices is the real contribution. Second, the new method RL[1]-OUPR* is presented as a BONE instance but actually leaks the current observation into the prior, and two of the classification experiments lack repeated trials. Read the empirical claims with more caution than the framework claims.\n\nWhat's new and good: the five-axis decomposition (measurement model, auxiliary variable, conditional prior, posterior approximation, auxiliary weighting) is clean, and Table 3 convincingly places Kalman filtering, BOCD, continual learning methods, and bandits into that space. Section 2's math is standard but correct, and the paper is explicitly honest about what BONE does not cover. They ship a JAX library, which is real evidence of reproducibility.\n\nThe main soft spot is definitional. Table 1 defines the M.3 prior as a function of D_{1:t-1} only. But Algorithm 4 computes the runlength posterior using the current y_t (lines 2–4), uses that posterior to set the prior mean and covariance in Eqs. (19)–(20), and then updates θ_t with the same y_t as the likelihood in line 16. So y_t enters twice: once to choose the prior, once to update from it. That makes RL[1]-OUPR* not actually an instance of BONE as formally defined. It could be defended as a data-driven prior heuristic, but the paper does not do that; it just calls it a new instance. This needs fixing either in the method or in the framework's definition.\n\nThe second issue: Sections 4.1.2 and 4.1.3 report single-run misclassification curves and claim \"significantly outperforms\" with no error bars or repeated seeds. The bandit and segmentation experiments do include repetitions, so this is not uniform, but the highlighted classification claims are under-supported. The ε threshold in Eqs. (19)–(20) also gets no sensitivity analysis; since the method's reset behavior hinges on it, that is a gap.\n\nThe core unification claim holds. The framework is an organizational contribution, not a new theory, but it is useful for anyone working in online learning, continual learning, or changepoint detection. The new method may survive the data-peeking fix, but the current experiments do not give clean evidence.\n\nRecommendation: send this to peer review. It deserves serious refereeing—the taxonomy is solid, the code is a plus, and the definitional and statistical issues are fixable in revision, not fatal. I would bring it to a reading group focused on non-stationary Bayesian inference.","headline":"BONE's taxonomy is a keeper; its new method has a data-peeking problem and thin stats in two key experiments.","tokens_in":34685,"tokens_out":5434,"would_cite":true,"duration_ms":50168,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62M20","68T05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A single framework rewrites non-stationary online learning as five interchangeable choices.","keywords":["online learning","non-stationary environments","Bayesian online changepoint detection","runlength","generalised Bayes","prequential forecasting","continual learning","state-space models"],"falsifier":"Run RL[1]-OUPR* on a synthetic stream whose changepoints are known and whose segments contain slow drift, sweeping $\\epsilon$ from 0 to 1. If no intermediate $\\epsilon$ beats both the always-Ornstein-Uhlenbeck and always-reset extremes, or if the best $\\epsilon$ shifts erratically across runs of the same task family, the claim that one threshold handles both gradual and abrupt change fails in that setting.","tokens_in":33624,"feed_emoji":"🧩","tokens_out":6523,"duration_ms":66060,"temperature":0.7,"pith_summary":"This paper claims that probabilistic online learning in changing environments reduces to five independent choices: what model generates measurements, what auxiliary variable tracks the changes, what prior over parameters that variable implies, how the parameter posterior is approximated, and how hypotheses about the auxiliary variable are weighted. It packages these five slots as the BONE framework and shows that a wide range of existing methods, from Kalman filters to Bayesian online changepoint detection to continual-learning and bandit algorithms, are instances of this structure. The paper also fills the slots with a new combination, RL[1]-OUPR*, which tracks a single runlength hypothesis and switches between gradual Ornstein-Uhlenbeck drift and a hard reset depending on the estimated probability that a changepoint has occurred. On electricity-load forecasting, drifting classification, bandits, and segmentation tasks, the new combination is reported to match or beat standard baselines while using only one retained hypothesis. If the claims hold, researchers get a common vocabulary for comparing, reusing, and extending methods whose connections were previously scattered across subfields.","feed_headline":"Five choices unify online learning in changing worlds","feed_subtitle":"Kalman filters, changepoint detectors, and continual learning are all instances of the same skeleton.","key_machinery":"The load-bearing object is the five-slot predictive decomposition given in equation (6) of the paper: measurement model $h$ with loss $\\ell$, auxiliary variable $\\psi_t$, conditional prior $\\pi$, parameter posterior $q$, and weighting function $\\nu$. The decomposition separates 'how the world changes' from 'how much to trust each hypothesis,' so swapping one slot yields a different algorithm while the update skeleton stays fixed. The paper's signature mechanism is the thresholded hybrid prior of equations (19)-(20): a convex Ornstein-Uhlenbeck combination of the previous belief and the initial prior when the posterior continuation probability is high, and a hard reset to the initial prior when that probability falls below $\\epsilon$.","core_discovery":"The central discovery is a decomposition of one-step-ahead prediction in a hierarchical state-space model into five named components. For parameters $\\theta_t$ and an auxiliary variable $\\psi_t$ that encodes non-stationarity, the predictive expectation factorizes as $\\hat{y}_{t+1}=\\sum_{\\psi_t}\\nu_t(\\psi_t)\\int h(\\theta_t,x_{t+1})q(\\theta_t;\\psi_t,D_{1:t})\\,d\\theta_t$, where $h$ is the measurement model, $\\nu_t$ weights hypotheses about the auxiliary variable, and $q$ is the parameter posterior. The paper argues that every known probabilistic online method is a specific choice of these slots plus the conditional prior $\\pi(\\theta_t;\\psi_t,D_{1:t-1})$, and it catalogs published filtering, segmentation, continual-learning, and bandit methods in this taxonomy. The proposed instance, RL[1]-OUPR*, uses the runlength as the auxiliary variable, keeps only the single most likely runlength hypothesis, and defines the conditional prior as an Ornstein-Uhlenbeck pull back toward the initial prior when the continuation probability $\\nu_t(r^{(1)}_t)$ exceeds a threshold $\\epsilon$, and as a hard reset to the initial prior otherwise. This thresholded hybrid is the paper's answer to environments that change both gradually and abruptly.","pith_inferences":["Beyond the paper, the five-slot decomposition suggests a search space: one could treat the slots as hyperparameters and systematically search over auxiliary-variable types, priors, and thresholds on a fixed benchmark suite.","The threshold $\\epsilon$ in the new method is the paper's most brittle hand-set hyperparameter, and nothing in the paper prevents learning it online with a second auxiliary variable, though the authors do not pursue this.","The framework implies a clean ablation discipline: to isolate the benefit of modelling gradual change, one compares methods that differ only in the prior slot, which the experiments begin to do.","The reported results hint that the right auxiliary-variable complexity is task-dependent, since RL[K]-PR overtakes the single-hypothesis method on some tasks when more hypotheses are kept; identifying when one hypothesis suffices is a natural next question."],"forward_implications":["A Kalman filter and a changepoint detector can be combined or compared by swapping one slot, so methods developed for filtering become usable for forecasting and vice versa.","The new single-hypothesis method RL[1]-OUPR* gives a low-memory option that, in the paper's reported experiments, performs competitively on data with both gradual and abrupt changes.","Outlier-robust likelihoods can be dropped into the framework without changing the rest of the algorithm, as demonstrated by the WoLF+RL[inf]-PR variant.","The framework exposes a memory-accuracy tradeoff: exact weighting over a growing auxiliary space is expensive, while DA[K] with K retained hypotheses bounds the cost at the price of pruning the hypothesis set.","Because the auxiliary variable is a free design choice, a method built for segmentation with runlengths can be transplanted to bandits or continual learning, which the paper demonstrates experimentally."],"supporting_citations":[{"why":"Supplies the runlength auxiliary variable and the recursive Bayesian online changepoint detection update that anchors RL-PR and the BOCD family.","marker":"Adams & MacKay (2007)"},{"why":"Supplies the moment-matched prior reset and the dependent-segment autoregressive model used for RL-MMPR and the segmentation experiments.","marker":"Fearnhead & Liu (2011)"},{"why":"Supplies the linear-Gaussian state-space filtering update that serves as the base C-LSSM instance of the framework.","marker":"Kalman (1960)"},{"why":"Supplies the changepoint-probability auxiliary variable with Ornstein-Uhlenbeck prior used by the CPP-OU baseline.","marker":"Titsias et al. (2024)"},{"why":"Supplies the Ornstein-Uhlenbeck conditional prior for continual learning that the new hybrid prior extends.","marker":"Kurle et al. (2019)"},{"why":"Provides the generalised Bayesian justification for replacing the likelihood with $\\exp(-\\ell)$, which the framework's measurement-model slot permits.","marker":"Bissiri et al. (2016)"},{"why":"Supplies the runlength-with-prior-reset construction for Thompson-sampling bandits that motivates the bandit experiments.","marker":"Mellor & Shapiro (2013)"},{"why":"Provides the electricity-load dataset with a known pandemic-related changepoint used in the hour-ahead forecasting experiment.","marker":"Farrokhabadi et al. (2022)"},{"why":"Supplies the WoLF-IMQ robust likelihood used in the heavy-tailed regression experiment.","marker":"Duran-Martin et al. (2024)"}],"fun_headline_variants":["Five choices explain every online learning method","BONE: the skeleton behind online learning in drift","One framework, five slots, all non-stationary learning","A unifying lens for changing-world learning","The modular recipe for online learning in flux"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The new method works only if the fixed threshold $\\epsilon$ and the single retained runlength hypothesis can be set so that continuation-versus-reset decisions are right in both gradual and abrupt phases, and the paper gives no principled rule for choosing $\\epsilon$.","fun_headline_variants_meta":{"raw":{"variants":["Five choices explain every online learning method","BONE: the skeleton behind online learning in drift","One framework, five slots, all non-stationary learning","A unifying lens for changing-world learning","The modular recipe for online learning in flux"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000719,"raw_usage":{"total_tokens":3292,"prompt_tokens":1072,"completion_tokens":2220,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":688,"completion_tokens_details":{"reasoning_tokens":2160}},"tokens_in":688,"tokens_out":2220,"duration_ms":15895,"temperature":1.0,"reasoning_tokens":2160,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:54:43.224519+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run RL[1]-OUPR* on a synthetic stream whose changepoints are known and whose segments contain slow drift, sweeping $\\epsilon$ from 0 to 1. If no intermediate $\\epsilon$ beats both the always-Ornstein-Uhlenbeck and always-reset extremes, or if the best $\\epsilon$ shifts erratically across runs of the same task family, the claim that one threshold handles both gradual and abrupt change fails in that setting.","supporting_citations":[{"cited_title":"Efficient bayesian analysis of multiple changepoint models with dependence across segments","cited_arxiv_id":null,"evidence_quote":"Supplies the moment-matched prior reset and the dependent-segment autoregressive model used for RL-MMPR and the segmentation experiments."},{"cited_title":"Continual learning with bayesian neural networks for non-stationary data","cited_arxiv_id":null,"evidence_quote":"Supplies the Ornstein-Uhlenbeck conditional prior for continual learning that the new hybrid prior extends."},{"cited_title":"Shestpaloff, Leandro S \\'a nchez-Betancourt, Jeremias Knoblauch, Matt Jones, Briol Fran c ois-Xavier, and Kevin P","cited_arxiv_id":null,"evidence_quote":"Supplies the WoLF-IMQ robust likelihood used in the heavy-tailed regression experiment."}],"review_version":1}