{"id":"b87d626b-0972-42e2-ad25-b3b80321bf9f","arxiv_id":"2412.11289","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Continual learning agents trained with CLEAR and EWC outperform deep learning baselines on non-stationary bug localization across several Apache Java projects, with lower training cost.","lead":"This paper tests whether continual learning agents (CLEAR and EWC) can keep locating bug-related code changes as software evolves, comparing them with deep learning baselines on seven Java projects. It reports that the continual learning agents match or beat the baselines on several ranking metrics in non-stationary settings while training much faster.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The CL-vs-DL comparison lacks a retrained or incrementally-updated baseline, so the reported non-stationary advantage may only reflect that CL agents were trained on the evaluation distribution while FLIM and RLOCATOR were not.","rationale":"The reader's weakest_assumption identifies precisely the most load-bearing issue: the baselines FLIM and RLOCATOR are never retrained or incrementally updated on the non-stationary data, while the CL agents are trained on that data. This makes the headline comparison between CL and DL techniques an apples-to-oranges comparison. The paper's motivating discussion explicitly frames retraining and incremental updating as the costly alternatives to CL, so omitting them as baselines is not a minor gap; it is the missing control needed to attribute the observed advantage to continual learning rather than to training-set overlap. The computational-effort claim is similarly affected because it compares one-time baseline training against CL training over the whole task sequence, not against the cost of maintaining a baseline under drift. I agree with the reader that the appropriate disposition is conditional: the empirical setup should be re-run with adapted baselines, and the regression-based reward enhancement should also be checked for temporal leakage of bug-fix information, before the central claim is accepted. My stress-test does not move the reader's verdict, so I set verdict_should_be to UNCHANGED.","tokens_in":38468,"tokens_out":6557,"duration_ms":60409,"concrete_test":"Re-run RQ1/RQ2 with two additional adapted baselines: (a) RLOCATOR and FLIM retrained from scratch on the combined stationary+non-stationary training split, and (b) incremental-update versions that periodically fine-tune on the non-stationary stream up to the bug-report date. Use the same 60:40 temporal split, five runs per configuration, and report MRR, MAP, top@1/5/10 on the non-stationary test set, plus total wall-clock training time for every method. If either adapted baseline reaches parity or exceeds CLEAR/EWC, the central claim fails; if CL agents still win under this missing control, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparison in RQ2 (Section 5.2) is confounded by a training-distribution asymmetry. The CL agents are trained cyclically on both stationary and non-stationary data, and the non-stationary data used for evaluation is the same stream they were trained on. In contrast, FLIM and RLOCATOR are trained only on changeset-files collected when bugs are fixed (Section 5.1, RQ1 method) and are then scored on non-stationary data with no adaptation. The paper's own introduction lists retraining and incremental updating as the standard remedies for concept drift, but neither is included as a baseline. Thus the reported gains (e.g., 'up to 61%' in MRR) do not establish that the CL mechanisms (replay/EWC) are responsible; they may simply reflect that CL agents saw the evaluation distribution during training while the baselines did not. The '5x less computational effort' claim has the same problem: it compares one-time baseline training against CL training on the full task sequence, not against the cost of retraining or incrementally updating the baseline on the non-stationary stream.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes applying two continual-learning agents, CLEAR (experience replay) and EWC (elastic weight consolidation), to bug localization over changeset-files and hunks, and augments their reward function with a logistic-regression model of bug-inducing factors. The authors evaluate the agents on seven open-source Java projects against three DL baselines (FLIM, RLOCATOR, FBL-BERT), reporting MRR, MAP, top@1/5/10, forgetting, and training time. The headline claims are that CL agents outperform DL baselines in non-stationary settings by up to 61% in MRR (and correspondingly in other metrics), reduce catastrophic forgetting, and require up to 5x less training effort.","tokens_in":38663,"tokens_out":12684,"duration_ms":103501,"significance":"Timely and practically motivated: concept drift is a real threat to DL-based bug localization, and the idea of adapting CL/DRL agents to ranking buggy changesets is interesting. The paper covers two distinct CL mechanisms, two granularities, seven projects, and provides a replication package; it also reports a forgetting metric and repeated runs. If the empirical claims were established under a fair comparison, the result would be a useful step toward adaptive bug localization. At present, however, the central comparison is confounded by a training-distribution asymmetry, and the regression-based reward appears to use oracle information from the fix commit. The contribution is therefore not yet demonstrated.","major_comments":[{"comment":"The RQ2 comparison is confounded by a training-distribution asymmetry. Section 5.1 states that 'FLIM and RLOCATOR baseline studies trained and evaluated their proposed approaches on changeset-files collected when bugs are fixed only,' while the CL agents are 'sequentially trained on both stationary and non-stationary data' (Section 5.2 Method). The CL agents are therefore exposed at training time to the same kind of non-stationary stream on which they are later evaluated, whereas the DL baselines are not. The paper itself (Section 1) identifies retraining and incremental updating as the standard remedies for concept drift, yet neither is included as a baseline. As a result, the headline improvements in the abstract and Finding 2 (up to 61% MRR, 44% MAP, etc.) may simply reflect different training distributions rather than the replay/EWC mechanisms. The same issue affects the FBL-BERT comparisons in Tables 9–10, where CL scores are averaged over stationary and non-stationary data but no retrained or incrementally updated FBL-BERT variant is supplied. Please add retrained/incremental baselines, or at minimum evaluate all techniques under identical training-data conditions.","section":"§5.1–5.2 (Tables 3–10)"},{"comment":"The logistic-regression reward appears to leak the bug-fixing oracle. The retained features Churn and PRE are defined using the fix commit: PRE is computed 'based on the set of files updated when fixing a given bug,' and Churn uses 'the diff associated with the bug commit under analysis' (Section 4.1.3, Table 1). Adding the fitted model's output to the reward in Eq. (5) thus rewards actions according to properties of the very commit that the technique is supposed to predict. Even if the logistic regression is fit only on training bugs, applying it to a test bug requires computing these features from that test bug's fix diff, which is the localization target. This likely inflates the '+Reg' results in RQ3 (Finding 3) and the Section 5.4 ablation, and it also affects the RLOCATOR+Reg variants. The features must be redefined using only information available before the fix (e.g., the latest pre-fix version), or the regression-enhanced results should be presented as an oracle upper bound.","section":"§4.1.3, Table 1, Eq. (5)"},{"comment":"The 'up to 5x less computational effort' claim is not an apples-to-apples comparison. Table 19 reports one-time training costs for FLIM/RLOCATOR on stationary fix-time data, while the CL agents are trained cyclically on both stationary and non-stationary tasks (Section 4.1). The cost of retraining or incrementally updating the DL baselines on the non-stationary stream—the standard remedies the paper motivates in Section 1—is not measured. Without that reference point, the training-time advantage cannot be attributed to continual learning rather than to the different training protocol, and the abstract's '5x less' claim is not established.","section":"§5.2, Tables 19–20"}],"minor_comments":[{"comment":"The text reports a performance drop of '9% to 194%' in top@1, top@5, and MRR, but for metrics bounded by [0,1] a relative decrease cannot exceed 100%; please clarify the calculation or the reference value.","section":"§3, Figure 1"},{"comment":"The list of Apache projects is given as '(AspectJ, Birt, JDT, PDE, Eclipse, Zxing, Tomcat, and Birt)', with Birt listed twice; since JDT is excluded, the resulting seven projects should be named consistently with Table 2.","section":"§4.2"},{"comment":"The text appears to cite 'FLIM [13] and RLOCATOR [37]', but in the reference list [13] is RLocator and [37] is FLIM; the citations should be swapped.","section":"§5.1"},{"comment":"Several captions repeat 'In bold are the best average performances' (e.g., Table 13); please remove the duplicate phrase.","section":"Tables 13–22"},{"comment":"Equation (2) and the surrounding text define Fisher information in terms of parameters learned on task A; the phrase 'the learned parameters θ*_A of a task B' should say 'of task A'.","section":"§4.1.2"},{"comment":"The paper repeatedly states that differences are statistically significant, but the tables do not report p-values or effect sizes; a compact significance summary should be included in the paper itself, not only in the replication package.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"Given the two load-bearing issues (training asymmetry and oracle leakage), the paper needs substantial additional experiments, but the underlying data and infrastructure seem reusable. A resubmission that adds retrained/incremental baselines and removes the fix-commit features from the reward could plausibly be accepted. I would not reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline result here—CL agents beating DL baselines by up to 61% in MRR on non-stationary data—doesn't hold up as stated, because the baselines were never trained on the non-stationary distribution. FLIM and RLOCATOR are trained on fix-time data only (Section 4.2, RQ1 method), then scored on non-stationary data the CL agents were explicitly trained on. That's not a fair test of whether the CL mechanisms (replay/EWC) help; it's mostly comparing models that saw the evaluation distribution against models that didn't. The paper itself, in the introduction, lists retraining and incremental updating as the standard remedies for concept drift, but neither is included as a baseline. Without those, the reported CL advantage could be an artifact of training on the test stream.\n\nThere's also a likely leak in the logistic regression reward. The features (Churn, PRE) are computed from the bug-fixing commit—the ground truth used as the oracle. Feeding that probability into the reward tells the agent which files were changed in the fix. That inflates the +Reg results and the ablation findings.\n\nWhat's genuinely useful: the paper applies two established CL methods (CLEAR, EWC) to bug localization for the first time, with a replication package, three baselines, and extensive ablations across seven projects. The claim that CL agents degrade less on non-stationary data than static models is plausible and worth testing properly. The computational-effort data are also interesting, though the '5x less' claim compares one-time baseline training against CL training on the full sequence, not against the cost of retraining baselines.\n\nI'd send this to peer review, but with conditions: the authors need to add retrained and incrementally-updated baseline variants, and address the oracle leakage in the reward. The core idea isn't nonsense; the evidence just doesn't support the headline yet. If a reviewer gets the setup fixed, the comparison could be a decent empirical contribution.","headline":"CL bug localization beats DL baselines only because the baselines never see non-stationary data; retrained baselines are the missing control.","tokens_in":39240,"tokens_out":1967,"would_cite":false,"duration_ms":17972,"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":"Continual learning agents localize buggy changesets more accurately than deep learning baselines when software code evolves, while training up to five times faster.","keywords":["Continual learning","Bug localization","Deep reinforcement learning","Concept drift","Catastrophic forgetting","Changesets","Bug-inducing factors","Software maintenance"],"falsifier":"Retrain FLIM, RLOCATOR, and FBL-BERT on the same non-stationary training stream, the commits between bug report and fix, given to CLEAR and EWC, and check whether their MRR, top@1, and top@5 on the non-stationary test split still trail by the reported margins.","tokens_in":38223,"feed_emoji":"🐞","tokens_out":7165,"duration_ms":58597,"temperature":0.7,"pith_summary":"The paper sets out to show that continual learning (CL) agents can keep locating buggy changesets accurately when software code evolves between the time a bug is reported and the time it is fixed, a situation the authors call non-stationary data. They train two deep reinforcement learning agents, CLEAR and EWC, cyclically on stationary and non-stationary changeset-files and hunks from seven Apache projects, and compare them with deep learning baselines. The reported result is that the CL agents beat the FLIM and RLOCATOR baselines in the non-stationary setting by up to 61 percent in MRR, 44 percent in MAP, 83 percent in top@1, 56 percent in top@5, and 66 percent in top@10, while needing up to five times less training time; against the BERT-based FBL-BERT baseline, they win on top@5 and top@10 but trail on MRR and top@1. Augmenting the reward function with a logistic regression score built from bug-inducing factors improves the CL agents on several projects and reduces forgetting. A sympathetic reader would take away that continual learning is a cheap way to keep bug localization useful as code drifts.","feed_headline":"Continual learning tops deep models on drifting bug data","feed_subtitle":"Trained on evolving code, the agents lift bug ranking by up to 83 percent and cut training effort fivefold.","key_machinery":"The central machinery is a deep reinforcement learning formulation of bug localization. The agent's state is a concatenation of CodeBERT embeddings for a bug report and for up to 31 changeset-files or hunks retrieved by an Elasticsearch index; its action is picking one item to move into a ranked list, and its reward is based on the rank of relevant changesets and the distance between them in the list. Continual learning comes from two mechanisms: CLEAR, a rehearsal method that mixes new and replayed experiences and corrects off-policy drift with V-Trace importance weights, and EWC, a regularization method that penalizes changes to weights important for previous tasks via the Fisher information matrix. A third component, the logistic regression model over bug-inducing factors, produces a scalar bug-probability that is added into the reward function.","core_discovery":"The central claim is that rehearsal- and regularization-based continual learning agents can handle concept drift in bug localization without the performance collapse seen in deep learning models trained on stationary, fix-time data. CLEAR, an experience-replay method that uses V-Trace off-policy correction, and EWC, a weight-consolidation method that uses the Fisher information matrix, are trained in sequence on stationary and non-stationary changeset-files and hunks. On non-stationary data the authors report large gains over FLIM and RLOCATOR, as summarized in the abstract, and they report that the agents mitigate catastrophic forgetting across the two tasks. They further report that adding a logistic regression score over bug-inducing factors, reduced to churn and pre-release bugs, improves performance on four of seven projects at the changeset-file level and three of seven at the hunk level, and lowers forgetting values. The paper also acknowledges that the BERT-based FBL-BERT baseline retains an advantage in MRR and top@1, while the CL agents lead on top@5 and top@10.","pith_inferences":["An implication the authors leave implicit is that the reported advantage is measured against baselines that were not retrained on the non-stationary stream; if FLIM, RLOCATOR, or FBL-BERT were periodically updated on the same evolving data, the gap could be materially smaller.","A testable extension would be to replace the logistic regression reward component with an online-learned bug-proneness estimator, so the prior knowledge adapts to the same drift the agent is learning from.","The CL agents' lower top@1 relative to FBL-BERT suggests a hybrid design that uses FBL-BERT's retrieval refinement for the first result and CL ranking for the full list could combine both strengths.","Because the evaluation covers seven Java projects, a natural next test is whether the same CL setup transfers across programming languages or across projects with different commit densities."],"forward_implications":["Bug localization tools built on CL agents can keep ranking relevant changesets as code evolves, without full retraining from scratch on every new version.","Developers can train CL agents for a new project with up to five times less compute than the deep learning baselines, making continuous updates practical on smaller infrastructure.","Adding cheap bug-inducing signals such as churn and pre-release bug history to the reward function improves ranking and reduces catastrophic forgetting, so prior project knowledge can be folded into the agent at no extra training-data cost.","Because CL agents mitigate forgetting across stationary and non-stationary tasks, the same mechanism can carry over to other software engineering tasks that suffer from concept drift, such as defect prediction, malware detection, and test-case prioritization, which the paper lists as recommended applications."],"supporting_citations":[{"why":"Supplies the DRL environment formulation and the RLOCATOR baseline that the CL agents are compared against.","marker":"[13]"},{"why":"Provides the FLIM baseline, a function-level interaction model for semantic bug localization.","marker":"[37]"},{"why":"Provides the FBL-BERT baseline at changeset-files and hunks levels, used for comparison.","marker":"[17]"},{"why":"Source of the CLEAR experience-replay continual learning algorithm.","marker":"[58]"},{"why":"Source of the elastic weight consolidation algorithm used by EWC.","marker":"[32]"},{"why":"Supplies the V-Trace off-policy correction algorithm used inside CLEAR.","marker":"[24]"},{"why":"Provides the EWC implementation and the forgetting metric used to evaluate the CL agents.","marker":"[55]"},{"why":"Provides the bug-inducing factors and logistic regression methodology used in the reward enhancement.","marker":"[65]"},{"why":"Supplies CodeBERT embeddings used to represent bug reports and changesets in the agent's observation.","marker":"[25]"}],"fun_headline_variants":["Continual learning beats deep models by 83% on buggy drift data","CL methods lift bug localization up to 83% on shifting code","Rehearsal-based CL outperforms deep nets for bug hunting on drift","On evolving datasets, continual learning boosts bug ranking by 83%","Continual learning for bugs: up to 83% gain, 5x less compute"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes that the correct baselines are deep learning models trained only on stationary, fix-time data; if those baselines were retrained or incrementally updated on the non-stationary data the CL agents receive, the reported advantage might shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Continual learning beats deep models by 83% on buggy drift data","CL methods lift bug localization up to 83% on shifting code","Rehearsal-based CL outperforms deep nets for bug hunting on drift","On evolving datasets, continual learning boosts bug ranking by 83%","Continual learning for bugs: up to 83% gain, 5x less compute"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000977,"raw_usage":{"total_tokens":4214,"prompt_tokens":1076,"completion_tokens":3138,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":692,"completion_tokens_details":{"reasoning_tokens":3038}},"tokens_in":692,"tokens_out":3138,"duration_ms":19946,"temperature":1.0,"reasoning_tokens":3038,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:05:36.630188+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain FLIM, RLOCATOR, and FBL-BERT on the same non-stationary training stream, the commits between bug report and fix, given to CLEAR and EWC, and check whether their MRR, top@1, and top@5 on the non-stationary test split still trail by the reported margins.","supporting_citations":[{"cited_title":"RLocator: Reinforcement Learning for Bug Localization","cited_arxiv_id":"2305.05586","evidence_quote":"Supplies the DRL environment formulation and the RLOCATOR baseline that the CL agents are compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the FLIM baseline, a function-level interaction model for semantic bug localization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the FBL-BERT baseline at changeset-files and hunks levels, used for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the CLEAR experience-replay continual learning algorithm."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the elastic weight consolidation algorithm used by EWC."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the V-Trace off-policy correction algorithm used inside CLEAR."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the EWC implementation and the forgetting metric used to evaluate the CL agents."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the bug-inducing factors and logistic regression methodology used in the reward enhancement."}],"review_version":1}