{"id":"8c09e7fd-b97c-42fc-a51d-9a29b2d3bcea","arxiv_id":"2501.05495","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"LSEBMCL replays EBM-generated pseudo-samples from previous tasks to reduce catastrophic forgetting, reporting top average scores on several NLP continual learning benchmarks.","lead":"A continual learning method for NLP uses a latent-space energy-based model to generate pseudo-examples of earlier tasks and replay them while training on new tasks. The paper reports state-of-the-art scores on three benchmark suites with a Mistral 7B backbone, but releases no code and leaves key sampling details unspecified.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The EBM replay loop is never specified: Eq. (14) samples latent z, but no equation maps z to a previous-task replay pair (x,y), so the reported gains cannot be attributed to the EBM.","rationale":"I read the full text and focused on whether the claimed mechanism, EBM-based replay, is actually defined. The reader's weakest assumption concerns the quality of EBM-generated pseudo-samples. I find a more fundamental defect: the manuscript never specifies how replay samples are produced. Equations (2)-(6) define an energy over input/label pairs (x,y), while Eqs. (7)-(15) define an energy-based prior over a continuous latent z. No equation trains α in the prior, no equation connects E(z) to EΘ(x,y), and no decoding procedure maps z to a replay pair (x,y). The role of the inference network is also ambiguous: Eq. (2) says AΨ maps x to y, but the figure caption says inputs are processed to produce z. Even if sample quality were high, this would need to be demonstrated, but here the generation mechanism itself is absent. This internal inconsistency is independent of empirical concerns such as the missing development set, which the paper itself discloses in Table I. The stated absence of a development set is a further reason the reported test-set-based selection of γ prevents accepting the SOTA claim. Because the method is not specified well enough to reimplement or to attribute the empirical gains, I agree with the reader's REJECT verdict; my concern does not change that verdict, though it identifies a different load-bearing point than the reader's weakest assumption.","tokens_in":9249,"tokens_out":5429,"duration_ms":58013,"concrete_test":"Independently re-derive the full replay loop from Section III alone: write down the loss used to train α in Eq. (8), the loss used to train β in Eqs. (11)-(13), and the exact steps that take a Langevin sample z_K from Eq. (14) to a replay example (x,y) for the next task. If the derivation requires any extra assumption not stated in the paper (e.g., a latent-to-label mapping or a decoding rule for y), the method is unspecified; if the derivation succeeds, check whether the resulting samples are informative by ablating the EBM prior (replace z_K with z0 ~ p0 in Eq. (14)) and measuring the change in Table II scores.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim requires that the latent-space EBM actually produces replay examples of previous tasks. The paper does not specify how. Section III-D defines two different energy functions: EΘ(x,y) in Eqs. (5)-(6) is a sum of per-task conditional log-losses over labels, while Eq. (8) defines a prior energy E(z) over a continuous latent z. No equation links them, and no training objective is given for the parameters α of pα(z). Eq. (14) draws z from that prior and Eq. (15) draws from the posterior pβ(z|x), but no step decodes z into a previous-task pair (x,y). The generator pβ(x|z) in Eq. (10) could produce text, but the answer y and the QA formatting are never specified. The inference network in Eq. (2) is trained to map x to y, yet the Figure 1 caption describes it as producing encoded representations z. As written, the replay loop cannot be run; consequently the SOTA numbers in Tables III and IV cannot be credited to EBM-generated replay rather than to the Mistral initialization, the GEN token formatting, or the replay fraction γ.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes LSEBMCL, a continual learning method for natural language processing tasks that uses a latent-space energy-based model (EBM) as a generative replay mechanism. The method is built on a Mistral 7B backbone and casts all tasks as question answering. After each task, the EBM is said to generate pseudo-samples of previous tasks, which are replayed when training on new tasks. The paper reports experiments on three small datasets under permuted task orders, on five DecaNLP tasks, and on four text classification datasets, claiming state-of-the-art results across all setups and near-multitask upper-bound performance. The central claim is that the EBM-based replay prevents catastrophic forgetting and outperforms prior replay, regularization, and architecture-based baselines.","tokens_in":9521,"tokens_out":3502,"duration_ms":35064,"significance":"If the proposed mechanism were fully specified and the empirical results were reproducible, the work would be a meaningful contribution: it applies latent-space EBMs to continual learning for NLP, addresses a broad set of tasks in a unified QA format, and reports low replay ratios with strong average performance. The paper also compares against multiple baselines and examines robustness to task order in the small-scale experiment. However, the significance is conditional on the method being implementable as described and on the reported gains being statistically credible. The manuscript currently lacks a complete specification of the replay loop, provides no code, and selects hyperparameters on test sets, so the central empirical claim cannot be credited in its present form. The stress-test concern about the missing link between sampling the latent variable and generating replay pairs is substantiated by the text.","major_comments":[{"comment":"The generative replay mechanism is not specified. Section III-D states that after training each task, the EBM generates samples based on data from previous tasks, and Eq. (14) draws z from the EBM prior pα(z) via short-run Langevin dynamics. However, no equation or procedure maps z to a previous-task replay pair (x, y). The decoder pβ(x|z) appears in Eq. (10), but it is never connected to the conditional distribution p(ym|xm) used in Eq. (6), and the answer component y is not generated at all. Moreover, no training objective is given for the prior parameters α: Eq. (11) trains β by maximum likelihood, but the parameters of pα(z) are never optimized. As written, the replay loop cannot be run, so the gains reported in Tables II-IV cannot be attributed to EBM-generated replay rather than to the Mistral initialization, the QA formatting, or the replay fraction γ.","section":"III-D, Eqs. (14)-(15)"},{"comment":"The energy function is internally inconsistent. Eq. (4) defines em(x, y) = −o2(zm) log pΘ(·|o1(zm), xm), while Eq. (6) defines em(x, y) = −log p(ym|xm). It is unclear which definition is actually minimized, how the softmax operators o1 and o2 are applied, and why both definitions appear in the same section. Eq. (5) also sums em over the task index m inside EΘ(x, y) for a single example, which conflates task identity with example identity. These ambiguities prevent a reader from reproducing the energy function or the training objective.","section":"III-B, III-C, III-D, Eqs. (4)-(6)"},{"comment":"The experimental evaluation selects hyperparameters on the test sets. Table I explicitly states, \"AS THIS WORK USES NO DEVELOPMENT SET, ONLY THE TRAINING AND TEST DATASETS ARE SHOWN.\" The results in Tables II-IV compare LSEBMCL across sampling ratios γ (0, 0.05, 0.2) and report the best configuration as the main result. Without a development set, there is no evidence that the reported γ values were not chosen based on test-set performance. Additionally, no confidence intervals or significance tests are provided, and Table IV reports averages over only two runs. The claim of state-of-the-art performance is therefore not statistically supported.","section":"IV-A, Table I; IV-B, Tables II-IV"},{"comment":"There is no ablation that isolates the contribution of the EBM replay. The paper's core premise is that latent-space EBMs produce pseudo-samples of previous tasks that prevent forgetting, but it never evaluates sample quality, shows example generations, or compares against LAMOL's language-model-generated replay at the same replay ratio. Without such an ablation, the reported improvements cannot be attributed to the energy-based prior rather than to other components such as the QA formatting, the choice of Mistral 7B as the backbone, or the increased replay fraction.","section":"III-D, IV-B"}],"minor_comments":[{"comment":"The notation in Eq. (8) is confusing: pα(z) is written as \"exp[Fα(z) − 1/2σ²∥z∥²] = exp E(z)\" without specifying the normalization constant in the equality, and the definition of E(z) is then given separately. Please make the proportionality explicit and define E(z) before writing the exponent.","section":"III-D, Eq. (8)"},{"comment":"Eq. (9) assumes x ∈ R^D and uses a Gaussian noise model, yet the text models are autoregressive over discrete tokens. The relationship between the continuous latent-space assumption and the discrete text generation model should be clarified.","section":"III-D, Eq. (9)"},{"comment":"The inference network in Eq. (2) is defined as AΨ(x) ≈ argmin_y EΘ(x, y), i.e., it predicts labels, but the Figure 1 caption describes it as generating encoded representations z. These two roles should be reconciled.","section":"III-A, Figure 1"},{"comment":"The column headers of Table II list all six task permutations but are not labeled as distinct orders, making the table difficult to interpret. Please add clear headers or a legend.","section":"IV-B, Table II"},{"comment":"The abstract and introduction claim \"state-of-the-art results in all experiments,\" but the comparisons are limited to specific baselines and the statistical support is weak. A more measured claim would be appropriate.","section":"Abstract and I"}],"recommendation":"reject","confidential_remarks":"The manuscript's central mechanism is underspecified to the point that the reported experiments cannot be reproduced from the text, and the evaluation methodology (no development set, test-set-based hyperparameter selection, no significance tests) would not support the stated claims even if the mechanism were clear. This goes beyond a local fix; the paper would need a complete methodological rewrite, new experiments with a proper development set, ablations of the replay component, and code release. I therefore recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper for one reason: it tries something genuinely new in NLP continual learning—using a latent-space energy-based prior as a replay generator—and reports numbers that, if real, would be strong (77.3 on five DecaNLP tasks vs. 78.2 multitask upper bound). The combination of Pang et al.'s latent EBM and LAMOL-style generative replay is a natural idea that, as far as I know, hasn't been tried for NLP CL before. That is the extent of the good news.\n\nThe method section does not hold up. The paper defines two different energy functions: EΘ(x,y) in Eqs. (5)–(6) as a sum of per-task conditional log-losses, and E(z) in Eq. (8) as an EBM prior over a continuous latent. No equation links them. Eq. (4) defines em(x,y) through operators o1 and o2, but Eq. (6) simply says em(x,y) = −log p(ym|xm). These are not the same thing, and the paper never says which is actually used. More importantly, the replay loop is never specified. Eq. (14) samples z from the EBM prior, and Eq. (15) samples z from the posterior pβ(z|x), but there is no equation mapping z to a previous-task replay pair (x,y). The generative model pβ(x|z) is introduced, but its training objective is given for pβ(x) alone, not for the joint energy EΘ(x,y). The abstract says the EBM \"samples data points from previous tasks,\" but as written, the pipeline cannot be run. The reported gains therefore cannot be attributed to the EBM, as opposed to the Mistral initialization, the GEN token formatting, or the replay fraction γ.\n\nEmpirically, the paper has other soft spots. Table I explicitly states there is no development set, and γ is selected and compared on test sets. There are no confidence intervals, no significance tests, and no code. The text classification results are averaged over two runs, but the DecaNLP table shows no variance at all. Given the method is underspecified, these numbers are not independently checkable. The decision to compare against eleven baselines is good, though; the baseline set is the right one, and if the method were properly specified, this would be a useful benchmark study.\n\nWho is this for? Someone who works on generative replay for NLP and is willing to reconstruct the missing pieces from context. The paper may contain a working idea, but it is not written clearly enough to evaluate. I would not cite it in its current form, but I would not write it off entirely either.\n\nRecommendation: send it to peer review, but with the expectation of major revision. The referees should demand a complete specification of the replay loop, a resolved energy definition, a development set, and code. Without those, the state-of-the-art claim is unsupported.","headline":"A plausible new combination of latent EBM and generative replay for NLP continual learning, but the method as written is not reproducible: the replay loop is unspecified, the energy equations are inconsistent, and hyperparameters are tuned on test sets.","tokens_in":10027,"tokens_out":1795,"would_cite":false,"duration_ms":19155,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Energy-based replay lifts continual NLP learning to 77.3","keywords":["continual learning","catastrophic forgetting","energy-based model","generative replay","latent space","large language models","task-incremental learning","pseudo-sample generation"],"falsifier":"Train LSEBMCL on the five DecaNLP tasks but replace the EBM-generated samples with samples generated by the base Mistral model itself at the same 5% ratio; if the average score does not drop meaningfully, the EBM prior is not the load-bearing component, and the reported gains could come from the QA format or replay fraction.","tokens_in":9038,"feed_emoji":"🧠","tokens_out":7205,"duration_ms":60130,"temperature":0.7,"pith_summary":"The paper proposes LSEBMCL, a continual learning method for NLP that generates pseudo-samples of previous tasks using a latent-space energy-based model (EBM) and replays them while training on a new task. The authors report that this approach largely prevents catastrophic forgetting: on a five-task DecaNLP sequence it reaches a mean score of 77.3, within 0.9 points of the multitask upper bound of 78.2, and on four text classification task orders it averages 80.2 with only a 5% sampling ratio. The method uses the Mistral 7B language model as backbone and a small MLP-parameterized EBM prior over latent vectors. If the claimed mechanism is sound, generative replay from a learned energy-based prior offers a way to retain old knowledge without storing raw examples.","feed_headline":"Energy-based replay lifts continual NLP learning to 77.3","feed_subtitle":"EBM-generated pseudo-data from earlier tasks keeps old skills alive with only 5% replay.","key_machinery":"The central mechanism is the latent-space EBM layer acting as an outer-generator. An energy-based model assigns low energy to likely inputs and high energy to unlikely ones; here it defines a prior $p_\\alpha(z)$ over dense latent vectors $z$ as an energy-based correction of a Gaussian, with the energy function given by $E(z)=F_\\alpha(z)-\\frac{1}{2\\sigma^2}\\|z\\|^2$, where $F_\\alpha$ is a small MLP. After each task, short-run Langevin dynamics (equation 14) draw latent samples from this prior, and a generative decoder maps them to pseudo-inputs of previous tasks, which are then replayed during training on the next task. This replay signal is what carries the anti-forgetting argument: old knowledge is preserved because the model keeps seeing generated versions of its earlier data.","core_discovery":"LSEBMCL claims that catastrophic forgetting in continual NLP learning can be substantially reduced by training an energy-based model in the latent space of the pretrained language model after each task, then drawing pseudo-samples of earlier tasks from this EBM prior via short-run Langevin dynamics and mixing them into the training data of the next task. The energy function is defined as a sum over tasks of negative log-likelihood terms, and the EBM prior takes the form $p_\\alpha(z) \\propto \\exp[F_\\alpha(z)]p_0(z)$, an energy-based correction of an isotropic Gaussian reference. The authors report state-of-the-art results across all experiments, including 77.3 average on five DecaNLP tasks (SQuAD 2.0, WikiSQL, SST, QA-SRL, WOZ) and 80.2 average accuracy on the four text classification orders, outperforming replay, regularization, and parameter-isolation baselines.","pith_inferences":["The paper does not ablate the EBM against the base language model's own generated samples at the same sampling ratio; a natural test is whether replacing the EBM with the LM's own generation changes the reported scores.","Because the experiments are task-incremental with known task boundaries, an open extension is whether LSEBMCL handles boundary-agnostic or class-incremental continual learning, where the EBM would need to be updated online.","One could test the memory claim directly by measuring how the EBM-generated samples' text quality changes as the task sequence grows, and whether quality predicts forgetting.","The latent-space EBM is trained per task; sharing a single continuously updated EBM across tasks could remove the need to keep per-task model copies."],"forward_implications":["Continual NLP models can be trained over long task sequences without storing raw user data, since the EBM prior and its parameters are the only retained trace of previous tasks.","The reported 5% sampling ratio reaching near-upper-bound performance suggests generative replay can be highly sample-efficient, reducing memory and compute relative to large exemplar buffers.","Because the method relies on no task-specific parameters, it can be applied to task-agnostic continual learning settings where task identity is not available during inference.","The same outer-generator EBM pattern could extend to other sequence modalities, such as speech or code, provided a generative decoder for latent vectors exists."],"supporting_citations":[{"why":"Supplies the latent-space energy-based prior formulation that LSEBMCL adapts as its sampling mechanism.","marker":"[2]"},{"why":"Prior use of EBMs in continual learning, which the paper contrasts by placing the EBM as an outer-generator rather than the core model.","marker":"[3]"},{"why":"LAMOL, the generative replay baseline that LSEBMCL extends and outperforms on DecaNLP tasks.","marker":"[12]"},{"why":"Mistral 7B, the pretrained backbone used by LSEBMCL and all compared methods.","marker":"[24]"},{"why":"DecaNLP, the source of the five tasks and the unified QA formatting used for continual training.","marker":"[25]"},{"why":"Source of the four text classification task orders and the MBPA++ baseline.","marker":"[26]"},{"why":"RVAE-LAMOL baseline used in the DecaNLP comparison.","marker":"[27]"},{"why":"HMI-LAMOL baseline used in the DecaNLP and text classification comparisons.","marker":"[28]"}],"fun_headline_variants":["Latent EBM pseudo-replay hits 77.3, besting continual NLP baselines","Energy-based memory in latent space cuts catastrophic forgetting to SOTA","EBM samples past tasks via Langevin, lifts continual NLP to 77.3","Short-run Langevin EBM replay sets new state-of-the-art in continual learning","Pseudo-data from latent EBM prevents forgetting, top marks on DecaNLP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that pseudo-samples drawn from the latent EBM prior faithfully represent previous tasks well enough that replaying them preserves old-task performance.","fun_headline_variants_meta":{"raw":{"variants":["Latent EBM pseudo-replay hits 77.3, besting continual NLP baselines","Energy-based memory in latent space cuts catastrophic forgetting to SOTA","EBM samples past tasks via Langevin, lifts continual NLP to 77.3","Short-run Langevin EBM replay sets new state-of-the-art in continual learning","Pseudo-data from latent EBM prevents forgetting, top marks on DecaNLP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1435,"prompt_tokens":896,"completion_tokens":539,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":432}},"tokens_in":512,"tokens_out":539,"duration_ms":5247,"temperature":1.0,"reasoning_tokens":432,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:12:47.159825+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train LSEBMCL on the five DecaNLP tasks but replace the EBM-generated samples with samples generated by the base Mistral model itself at the same 5% ratio; if the average score does not drop meaningfully, the EBM prior is not the load-bearing component, and the reported gains could come from the QA format or replay fraction.","supporting_citations":[{"cited_title":"Learning latent space energy-based prior model,","cited_arxiv_id":null,"evidence_quote":"Supplies the latent-space energy-based prior formulation that LSEBMCL adapts as its sampling mechanism."},{"cited_title":"Energy-based models for continual learning,","cited_arxiv_id":null,"evidence_quote":"Prior use of EBMs in continual learning, which the paper contrasts by placing the EBM as an outer-generator rather than the core model."},{"cited_title":"Episodic memory in lifelong language learning,","cited_arxiv_id":null,"evidence_quote":"Source of the four text classification task orders and the MBPA++ baseline."},{"cited_title":"Rvae-lamol: Residual variational autoencoder to enhance lifelong language learning,","cited_arxiv_id":null,"evidence_quote":"RVAE-LAMOL baseline used in the DecaNLP comparison."},{"cited_title":"Gener- ative replay inspired by hippocampal memory indexing for continual language learning,","cited_arxiv_id":null,"evidence_quote":"HMI-LAMOL baseline used in the DecaNLP and text classification comparisons."}],"review_version":1}