acceptodds
Under review as a conference paper at ICLR 2027

PulseLog: Context-Conditioned Behavioral Dynamics for Noise-Resilient Log Anomaly Detection

Abstract

Log anomaly detection requires inferring system behavior from observations. Their semantic and contextual evidence can be distorted by template variation, parameter errors, formatting changes, and other observation noise. Although existing methods have advanced semantic representation learning and temporal dependency modeling, they largely treat observed events as homogeneous evidence. Persistent operational regimes, transient deviations, and runtime context are entangled in a single latent trajectory. This prevents the detector from separating observation-induced perturbations from behaviorally meaningful changes. It also makes anomaly decisions sensitive to benign fluctuations and contextual shifts. We propose PulseLog, a context-conditioned behavioral dynamics model for noise-resilient log anomaly detection. PulseLog maps log observations to semantic anchors. It calibrates runtime context. It adaptively routes persistent and transient evidence at multiple temporal rates. It evolves a latent behavioral state through context-gated dynamics. A deviation probe identifies departures from context-specific state-transition patterns. This enables window-level anomaly detection based on deviations in system-state dynamics. Experiments across five public log systems demonstrate strong detection performance under standard evaluation, observation-noise perturbations, limited training data, and cross-system adaptation. PulseLog also has a compact model footprint. The results indicate that explicitly modeling context-conditioned behavioral dynamics improves robustness under the evaluated log observation conditions.

Primary area:
learning on time series and dynamical systems
Keywords:
Log anomaly detection, system-state modeling, robust representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_XRdxOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PulseLog, a supervised method for detecting anomalies in system logs. It uses log messages and runtime information to learn how system behavior changes over time. The model combines long-term patterns and short-term changes to identify unusual behavior. The authors test PulseLog on five public datasets and compare it with 15 baselines. It achieves the best or tied F1 scores, although the gains are small on most datasets. The paper also tests limited training data, cross-system transfer, and robustness to LLM-generated noise. The results are promising, but it remains unclear how well the method handles real-world log corruption.

Critical strengths
  1. The paper addresses an important problem in log anomaly detection. It aims to handle noisy log messages while capturing changes in system behavior over time.

  2. The method combines long-term patterns, short-term changes, and runtime context. This provides a useful way to model different types of information in system logs.

  3. The experiments cover five public datasets and 15 baselines. The authors also test noise robustness, limited training data, and cross-system transfer. These additional experiments help assess the method under different settings.

Critical weaknesses
  1. Log anomaly detection has been extensively studied, with existing methods already achieving strong results on standard benchmarks. The proposed method mainly combines established techniques, and its improvements over existing approaches are limited. It remains unclear what new challenge the method addresses or why the added complexity is necessary.

  2. The noise formulation is not well justified. The authors introduce several types of synthetic corruption and predefined noise rates, but do not explain how these choices relate to noise observed in real systems. It is also unclear whether the proposed method models noise explicitly or simply learns features that are less sensitive to the generated perturbations.

  3. The improvements over strong baselines are limited on most datasets. For example, the F1 gains are only 0.02 percentage points on HDFS v3 and 0.08 on BGL. These small differences make it difficult to justify the added model complexity. Several selected benchmarks are already close to saturation, with existing methods reaching nearly 99% or even 100% F1. These settings provide limited evidence that the proposed method can handle more difficult or realistic anomaly detection problems.

  4. The paper claims to detect deviations in system-state transitions. However, the model is trained using standard binary classification loss, without directly validating the learned transitions. It remains unclear whether the model actually captures meaningful behavioral changes or simply learns discriminative features.

Questions for authors
  1. What is the main methodological difference between PulseLog and existing log anomaly detection methods? Given the small performance gains on most benchmarks, what specific limitation does the proposed architecture address?

  2. How were the synthetic noise types and corruption rates selected? Is there evidence that they reflect noise observed in real production systems, beyond format and label consistency checks?

  3. How do the authors verify that the learned state transitions represent meaningful system behavior rather than features learned for binary classification?

  4. Since several benchmarks are already close to saturation, what evidence supports the method's advantages in more challenging anomaly detection settings?

1 comment

Thank you for your constructive feedback. We added production logging examples and noise tests on HDFS v1 and Thunderbird (Sec. 4.4), along with behavioral state and transition probing analysis (Sec. 4.5).

  1. The distinction lies in the behavioral representation and how it evolves. PulseLog constructs persistent and transient evidence separately, then uses runtime context to condition state retention, evidence injection, and the local residual update. The same semantic event can therefore lead to different state changes under different runtime conditions. This addresses the entanglement of stable operation, benign observation changes, and short-lived fault evidence in a shared event representation. Sections 3.2–3.4 specify this observation-to-trajectory dependency. We use established neural operators, but organize them around conditional state evolution rather than treating runtime parameters only as appended features. The added structure is intended to preserve useful transient evidence while reducing sensitivity to unreliable local observations. To evaluate this benefit, we have extended the noise robustness experiments to HDFS v1 and Thunderbird. The results show clear performance advantages under noise conditions encountered in practice, supporting the architectural design beyond small gains on already saturated benchmarks.

  2. We have expanded Sec. 4.4 and Appendix B.4 to document the production motivation and corruption protocol. LogRobust (FSE 2019) describes logging-statement evolution and processing noise, including evaluation on Microsoft’s online-service logs. AWS ALB documentation illustrates status-code pairs such as 200 200 for successful requests and 502 - for failures, as well as a target-processing time of -1 when a target fails to respond before timeout. Slack’s May 2020 incident report describes increased HTTP 503 errors triggering alerts. LogRobust motivates testing robustness to observation instability, while the AWS and Slack examples show that runtime fields can carry genuine fault evidence that must be preserved. Accordingly, Appendix B.4 specifies constrained perturbations to templates, tokens, parameters, and values, excluding edits that alter fault semantics. PulseLog learns context-conditioned behavioral representations from imperfect observations; it does not estimate an explicit probabilistic corruption process. Our robustness experiments evaluate whether this mechanism tolerates observation noise while preserving fault-relevant evidence.

  3. The learned states are designed to summarize behavioral history by retaining persistent operational context and transient evidence, with state evolution conditioned on runtime context. To make this interpretation testable, we added a frozen-representation diagnostic protocol in Sec. 4.5/Fig. 2. It probes execution states from frozen hidden states and transitions from adjacent hidden-state pairs, using current-event features, a random encoder, and pooled history as baselines. An adjacent-event control receives the same two events to assess the transition information retained by the representation. The detection and robustness results demonstrate that this context-conditioned state design enables effective anomaly detection under imperfect observations. The learned states retain substantially richer behavioral information than the baseline representations, supporting the effectiveness of our context-conditioned state design.

  4. Sec. 4.4/Table 6 now evaluates corruption on BGL, HDFS v1, and Thunderbird. At 5% noise, PulseLog exceeds the strongest baseline by 2.50 and 2.15 F1 points on HDFS v1 and Thunderbird, respectively. Table 4 separately varies training size from 5K to 20K windows with a fixed test set. PulseLog is best or tied best across budgets, reaching 97.74% F1 on Hadoop with 5K windows. Together with the ablations, these results demonstrate the architecture’s effectiveness under noise and limited training data.

Reviewer_Zc3sOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

PulseLog models log anomaly detection as departure from context-conditioned system-state dynamics rather than unusual messages alone. It maps observations to semantic anchors, calibrates runtime context, separates persistent and transient temporal evidence, and probes deviations from learned transitions (Figure 1; Section 3). PulseLog is a carefully motivated and broadly evaluated model, with useful efficiency and appendix analyses. The principal limitations are that some benchmark gains are near-ties, the noise-robustness experiment is controlled and narrow, and the transfer and deployment claims need more discriminating comparisons. These are mostly addressable through claim calibration and targeted protocol clarification.

Critical strengths
  1. A coherent behavioral modeling hypothesis.The persistent/transient distinction, context calibration, gated dynamics, and deviation probe correspond to recognizable failure modes of ordinary semantic or sequential detectors (Introduction; Figure 1; Section 3). This gives the architecture a clearer rationale than a generic stacked encoder.
  2. Broad evaluation and useful reporting depth. Five systems, numerous baselines, three random seeds, cross-system adaptation, training-size variation, architecture substitutions, and hyperparameter sensitivity are included (Table 1; Sections 4.3–4.5; Appendices B and F). The appendix also reports variability rather than presenting only a single best run.
  3. A meaningful efficiency profile. The model uses 1.20M parameters, 41.43M FLOPs, and 13.8MB reported model size on BGL (Table 2), making the deployment question more tangible than a purely accuracy-oriented comparison.
Critical weaknesses
  1. Several headline comparisons are too close to support a broad superiority claim. HDFS v3 F1 is 0.9368 versus 0.9366 for DeepLog; BGL is 0.9869 versus 0.9861 for LogDLR; Thunderbird reaches a ceiling near 1.000 (Table 1; Section 4.2; Appendix F). This is a claim-calibration issue. Report paired confidence intervals for every benchmark and emphasize settings with practically meaningful differences rather than counting near-ties as wins. 2.The robustness evidence is narrower than the title suggests. 3.Cross-system adaptation still uses substantial target supervision. 4.Metrics and windows may mask operational false alarms.
Questions for authors
  1. How much does source training improve cross-system F1 over a target-only model with exactly the same 20,000 labeled target windows?
  2. Were BGL noise tests performed with a clean-trained detector evaluated on noisy logs, or was each noise rate used in both training and testing? Please give the exact split protocol.
  3. What are the confidence intervals for the HDFS v3 and BGL differences, and what is the end-to-end online processing rate including parsing and embedding?
1 comment

Thank you for your constructive feedback. We added target-only adaptation comparisons (C.2), clarified the noisy training and testing protocol (Sec. 4.4 and B.4), and calculated approximate confidence intervals for the small clean-data differences.

  1. Appendix C.2 compares Pretrained + FT with Target-only across 100–20K target labels, using the same architecture, target windows, and evaluation splits. At 20K labels, pretraining improves F1 by 2.01 and 2.48 percentage points on HDFS v1 → BGL and Hadoop → BGL, respectively, showing a benefit from source initialization under identical target supervision.

  2. Both training and testing use the same corruption rate, with an 80:20 chronological split (Sec. 4.4; Appendix B.4). Noise is applied before embedding, preserving window boundaries and labels; all methods receive the same fixed corrupted partitions. Table 6 now extends this evaluation from BGL to HDFS v1 and Thunderbird.

  3. Mean F1 gains of 0.02 percentage points over DeepLog on HDFS v3 and 0.08 over LogDLR on BGL. Welch 95% intervals from the three-run summaries are [−0.197, 0.337] and [−0.034, 0.274] percentage points, respectively. PulseLog’s advantages are more pronounced under noise: at 5% corruption, it exceeds the strongest baselines by 1.75, 2.50, and 2.15 F1 percentage points on BGL, HDFS v1, and Thunderbird (Sec. 4.4/Table 6).

  4. PulseLog has 1.20M parameters, requires 41.43M FLOPs, and has a footprint of 13.8 MB. It achieves an encoding rate of approximately 949 messages/s and a processing throughput of approximately 800 messages/s, equivalent to 40 non-overlapping 20-event windows per second.

Reviewer_AGprOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies log anomaly detection under noisy observations, for example template changes, corrupted parameters, or format changes. The proposed method, PulseLog separates each message into its template and runtime parameters. It aggregates template information at two temporal scales and uses the runtime parameters to control how a latent state evolves. A learned readout then predicts whether a window is anomalous, using window-level labels for training. The theoretical analysis gives conditions for stable state updates and predictions under bounded perturbations, and discusses the bias–variance tradeoff of the two temporal scales. Experiments compare PulseLog against 15 baselines on five public datasets, reporting the best or tied-best F1 scores. The performance gains are clearer on HDFS-v1 and Hadoop, but small on HDFS-v3 and BGL, with no clear advantage on Thunderbird.

Critical strengths
  • The problem is well motivated and practically important. Real logs are noisy, and a useful detector should tolerate changes in how events are recorded while still catching real failures.
  • The method is reasonable. Using two temporal scales and letting the runtime parameters control the state update both make sense, and the single-scale comparison in Table 9 supports the design. The stability and bias–variance results in the appendix are simple but correct (This part is checked by AI).
  • The experiments are extensive. I appreciate the comparisons across five datasets and 15 baselines, with additional experiments on noise, limited training data, and cross-system adaptation.
Critical weaknesses

Major concerns

  • I find the technical novelty limited. The state update in Eq. (4), apart from the gated residual, is the discretized state-space update used in selective state-space models such as Mamba (Gu & Dao, 2023). In PulseLog, the main difference seems to be that the controller also receives the runtime context rtr_t. It is not clear to me how this differs from giving rtr_t to a standard selective state-space model as an additional input. Please explain what is technically new in these designs, and whether the Mamba backbone in Table 8 receives the same runtime-context input. If it does not, a comparison where it does would be helpful.
  • Key parts of the method are not clear. The Deviation Probe is central to the paper's claim that anomalies are detected as deviations in state transitions, but Section 3.4 only says it is a learned map from the state trajectory to a logit. I could not find what it computes or how it measures a "behavioral transition deviation". The same holds for the router, the maps, the controller, and the character-level encoder. Without these, the method cannot be reproduced.

Other Concerns

  • I wonder how useful Thunderbird dataset is. If it can be tested perfectly, it should be replaced by a more challenging dataset.
  • The benefit of cross-system pretraining is unclear. The transfer experiments fine-tune on 20K labeled target windows but do not compare against PulseLog trained only on those same windows.
  • PulseLog is trained with window labels, but several baselines are unsupervised or self-supervised, and the LLM baselines are zero-shot. It is not clear how each baseline was trained, and if the settings are comparable. Without this, it is hard to tell how much of the gap comes from the method and how much from the supervision.

On paper writing

  • Line 168, sentence-
  • I like Figure 1 but I think it is crowded and hard to read.
  • Table 2 has ICLR 2027 for the proposed method.
Questions for authors

Can the same idea help with anomaly detection in other noisy sequential data, or does it rely on properties specific to system logs?

1 comment

Thank you for your constructive feedback. We added component implementation details (A.7), the Mamba + Context baseline (C.3.1), target-only adaptation comparisons (C.2), and noise tests on HDFS v1 and Thunderbird (Sec. 4.4).

  1. Appendix A.7 now specifies the character encoder, signature maps, router, controller, and Deviation Probe. The probe computes D=b+∑twt⊤RMSNorm(ht)D=b+\sum_t w_t^\top\mathrm{RMSNorm}(h_t), a learned trajectory discriminant. Eq. (A.18) decomposes this score into an initial-state term and weighted increments of normalized states, explicitly showing how trajectory changes contribute to anomaly prediction.

  2. We added Mamba + Context in Appendix C.3.1/Table 8. PulseLog improves BGL F1 by 3.57 percentage points with only 3.1% more FLOPs. This supports the value of separating persistent and transient evidence and using context to control their routing and state updates, beyond simply supplying context to the backbone.

  3. Thunderbird represents a logging system and broadens system diversity. Its saturated clean results offer limited discrimination, so we added noisy Thunderbird and HDFS v1 evaluations in Sec. 4.4/Table 6. At 5% noise, PulseLog exceeds the strongest Thunderbird baseline by 2.15 F1 percentage points, making this dataset informative for evaluating noise robustness.

  4. Appendix C.2 adds the requested Target-only control with the same architecture, target windows, and evaluation splits. At 20K target labels, Pretrained + FT improves F1 by 2.01 and 2.48 percentage points on HDFS v1 → BGL and Hadoop → BGL, respectively, demonstrating a benefit from source pretraining under identical target supervision.

  5. Unsupervised baselines model normality, self-supervised baselines use prediction or reconstruction objectives, and LLMs use zero-shot prompts (B.3). Prior log anomaly detection studies likewise compare methods with different supervision requirements on shared benchmarks; we follow the same convention, using common test splits and metrics. The component controls in C.3–C.4 separately assess our architectural contributions.

Reviewer_snkwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper frames log anomaly detection as "context-conditioned behavioral state dynamics." PulseLog has five components: Semantic Anchor projects template semantics into a dynamics space; Context Calibrator generates context vectors from runtime parameters; Multi-Rate Evidence Router splits evidence into persistent and transient streams at different time scales; Context-Gated Dynamics updates the hidden state event by event, conditioned on both evidence streams and context (structurally close to selective state space models); and Deviation Probe reads the full state trajectory to produce a window-level anomaly probability. Experiments include standard comparisons on five LogHub datasets (15 baselines, including two LLMs), noise injection and training size experiments on 100K subsets, three cross-dataset transfer setups, ablations, backbone replacements, and stability and bias-variance analyses in the appendix.

Critical strengths

1.Defining anomalies at the level of state trajectory transition deviations, rather than individual log entries or events, is the right starting point. Anomalies in logs typically show up not in any single message but in transition patterns across events; the same template can carry different meanings under different runtime parameters. The persistent/transient decomposition also makes intuitive sense: steady-state operation calls for a long window, sudden shifts for a short one. In the ablation, removing the transient stream causes the largest F1 drop (around 3 percentage points), confirming this component actually matters. 2.The experimental coverage is unusually broad for this field. The 15 baselines include GPT-4o and DeepSeek-Flash; training size and noise are varied independently on 100K subsets with a fixed 2.7% class ratio; cross-dataset transfer is tested in three setups; and backbone, fusion method, and trajectory readout each get their own control group. This gives readers a full picture of the model's behavior across conditions, rather than a single main table to go by.

Critical weaknesses
  1. The incremental attribution of the core contribution is insufficient. The introduction and Section 3.1 frame "runtime context directly modulating state transitions" as the key distinction from prior work, yet the Section 4.5 ablation only removes transient evidence, Multi-Rate Fusion, the residual path, and the Deviation Probe without isolating runtime context. Table 8 in Appendix C.1 replaces the dynamics backbone entirely and does not specify which context input paths each alternative retains, so the performance gap cannot be attributed to context conditioning alone either. Appendix A does analyze context to some extent: after Theorem 1, the paper notes that context can shift the memory timescale via the step size, and A.6 incorporates gate perturbations into the robustness bound. These are stability conclusions under a given model with bounded-perturbation assumptions, though, and cannot answer "how much does adding context improve over not adding it, and under what conditions." The issue is not a lack of theoretical discussion but the absence of a direct context-free empirical ablation.
  2. The advantage on the main table is highly concentrated. PulseLog leads DeepLog by 0.02 points on HDFS v3, LogDLR by 0.08 on BGL, and five methods all reach 100% (to four decimal places) on Thunderbird. The clear gap comes mainly from HDFS v1 (F1 90.09% vs. 87.43%), where PulseLog has Precision 84.57% and Recall 96.37%, the opposite direction from most strong baselines that show high Precision and lower Recall. Appendix B.2 states that PulseLog uses a fixed 0.5 threshold, so the issue is not that the authors tuned their threshold; rather, the paper only reports P/R/F1 at this single operating point without a PR curve or ranking metric, leaving readers unable to tell whether the 90.09% advantage reflects overall ranking quality or whether this particular operating point just happens to be favorable. The title's "noise-resilient" claim also rests mainly on a synthetic noise experiment on BGL alone, where PulseLog leads Lograph by only 0.44 points at 1% noise; generalizing a robustness claim from a single dataset is a stretch.
Questions for authors
  1. Please add an ablation that removes only runtime context by setting r_t to a constant throughout: in the Eq. (2) weights, the Eq. (3) signatures, the subsequent router ξ_t = ψΘ(p_t, q_t, r_t), and the Eq. (4) controller whose input is (h_t^(ℓ−1), r_t), while keeping all other structures intact. Fixing the abstract control variable in Eq. (1) alone is not sufficient to rule out context entering the model through other paths. If this ablation barely lowers F1, the core motivation for context conditioning in Section 3.1 needs to be narrowed.
  2. Appendix B.2 states that PulseLog uses a fixed 0.5 probability threshold and that baselines fit method-specific thresholds on the training set before freezing them. Please provide the AUPRC/PR curve on HDFS v1, calibration information, and how P/R/F1 change under other thresholds, to clarify whether the 90.09% F1 reflects overall ranking quality rather than this particular operating point being favorable. Also please discuss whether a Precision around 85% is acceptable in the target operational setting.
  3. Section 4.3 labels the cross-dataset experiment "Few-Shot Cross-Dataset Adaptation," but Table 3 actually fine-tunes on 20K labeled samples from the target system, which is far from the usual meaning of "few-shot." Please explain why 20K was chosen and provide transfer results as the target label budget varies from hundreds to thousands to 20K. If the advantage narrows substantially at smaller budgets, the method primarily addresses cross-system adaptation rather than few-shot transfer, and the framing in the title and conclusions should be adjusted accordingly.
  4. The noise experiment is conducted on BGL only. Your method uses Drain for template parsing and a character-level encoder for parameters, which should theoretically help with template- and parameter-level perturbations. Does this advantage hold on other datasets? Conversely, given that Precision on HDFS v1 is already low, what happens when noise is added there?
1 comment

Thank you for your constructive feedback. We added a controller context ablation (C.4), precision–recall analysis (C.1), threshold sensitivity tests (E.4), and target-only adaptation comparisons (C.2), and extended noise evaluation to HDFS v1 and Thunderbird.

  1. Appendix C.4 (Table 11) shows that removing the controller’s direct context input reduces BGL F1 by 7.25 percentage points at 2% noise, even with context retained upstream. This isolates the contribution of direct context conditioning to noise robustness.

  2. Appendix C.1 adds PR curves under clean and noisy conditions. Appendix E.4 shows that F1 varies by only 0.55 percentage points across thresholds 0.3–0.6, with the default threshold 0.04 points below the maximum. On HDFS v1, approximately 85% precision accompanies 96% recall; deployment thresholds should reflect the alert budget.

  3. Appendix C.2 compares Pretrained + FT with Target-only across 100–20K target labels, holding the architecture, target windows, and evaluation splits fixed. At 20K labels, pretraining improves F1 by 2.01 and 2.48 percentage points on HDFS v1 → BGL and Hadoop → BGL, respectively, showing benefits beyond target supervision alone.

  4. Section 4.4/Table 6 adds HDFS v1 and Thunderbird alongside BGL at 1%, 2%, and 5% noise. PulseLog achieves the highest F1 across all nine conditions, exceeding the strongest baseline by 2.50 percentage points on HDFS v1 at 5% noise. These results extend the robustness evidence beyond BGL, with matched corruption rates in training and testing.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.