acceptodds
Under review as a conference paper at ICLR 2027

ECLIPSE: An Executable Contextual Intervention and Planning Agent for Time-Series Tasks

Abstract

Real-world time series cannot be fully characterized by historical observations alone, because their dynamics also depend on sparse contextual clues, external events, and domain knowledge. Existing methods often flatten such context into a prompt and require a single model pass to perform both semantic grounding and numerical extrapolation, leading to brittle reasoning. We propose ECLIPSE, an agentic system that recasts context-rich time-series reasoning as executable scenario reconstruction. ECLIPSE assembles a task-relevant scenario state from textual clues, retrieved temporal precedents, and domain priors, then compiles it into a temporal program. A unified planner selects contextual constraints and temporal operations, while foundation models and executable operators perform numerical computation. An audit agent checks contextual and numerical consistency, triggering replanning or a fallback to the unedited forecast when necessary. Across context-rich forecasting and temporal QA benchmarks, ECLIPSE outperforms direct prompting with the same language model, and a controlled comparison under a shared toolchain attributes the gain to the orchestration of history-side and future-side interventions rather than to tool access alone.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Time-series reasoning, LLM agents, Time-series forecasting, Temporal question answering

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 3 reviews, above 27% of the others.
Reviewer_971oOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ECLIPSE, an LLM-based agent framework for time-series forecasting and question answering with textual context. Instead of using an LLM to generate forecasts directly, the method uses it to plan a sequence of operations on the time series. The system first determines the task type and whether additional contextual knowledge is needed. For context-dependent tasks, it uses scenario-specific rules and retrieves reference patterns from a Temporal Archetype Bank. The planner can then modify the observed history before forecasting with Moirai and adjust the resulting forecast according to future events or constraints. For QA tasks, it instead generates a plan to extract the relevant temporal evidence. An audit module checks the generated operations and outputs, with replanning or fallback when needed. The authors evaluate ECLIPSE on CiK and TemporalBench. On CiK, ECLIPSE with DeepSeek-Chat achieves an average RCRPS of 0.174, compared with 0.188 for TimeClaw, although the previously reported Llama-3.1-405B direct-prompting result is better at 0.159. ECLIPSE also achieves the best contextual reasoning accuracy across the four TemporalBench datasets. The paper further includes a shared-toolchain comparison and component ablations to examine the contribution of the different intervention steps.

Critical strengths

1.I find the controlled comparison in Table 3 useful. The authors keep the LLM, forecasting backbone, retrieved references, and tool library fixed, making it easier to assess the contribution of the proposed workflow rather than differences in the underlying models or tools. This is a meaningful addition to the evaluation, especially for an agent-based framework.

2.The paper is fairly transparent about its limitations. The authors acknowledge that ECLIPSE does not outperform the previously reported Llama-3.1-405B result on CiK or the standalone Moirai model on historical-information tasks. They also report the extreme errors in some ablations and the relatively low repair rates of the audit module. I appreciate that these unfavorable results are not omitted.

Critical weaknesses

1.Limited novelty and unclear attribution of the performance gains. My main concern is that the methodological novelty is rather limited. Most components, including LLM planning, retrieval, temporal preprocessing/post-processing, and verification, are already well established. The two-sided intervention design is reasonable, but the contribution seems to be mainly in how these components are combined. More importantly, Table 3 does not control for inference budget: ECLIPSE uses 7.2 LLM calls per task on average, compared with one for the single-agent baseline. It is therefore unclear how much of the gain comes from the proposed orchestration itself. Much of the supporting analysis also relies on the authors' DTR metric rather than the original RCRPS metric.

2.Concerns about robustness and generalization. The ablation results in Table 13 are concerning. Removing certain components leads to extremely large errors on a few instances, suggesting that the planner can generate numerically unstable interventions. The trimmed statistics help characterize typical performance but do not fully address this issue, particularly given the audit module's limited ability to repair unsupported interventions (Table 18). Also, the seven scenario categories are designed around CiK-style tasks, so it is unclear how well the framework generalizes to unseen types of context.

Questions for authors

1.Table 3 controls for models and tools but not inference budget (7.2 vs. 1 LLM calls per task). Could the authors provide a budget-matched comparison, even on a subset? It would also be helpful to see whether the improvement holds under RCRPS rather than DTR.

2.The baselines in Table 2 appear to be taken from the original TemporalBench paper. Were the dataset version and evaluation protocol identical? Also, what explains the large improvement on Causal Chambers-HU (68.7% vs. 28.67% for the strongest baseline), where external context is not required?

3.Table 13 shows catastrophic errors when certain components are removed. Does the executor have general safeguards against numerically unstable interventions? Would simple bounds on intervention magnitudes prevent these failures while preserving the gains in Table 4?

4.Could the authors clarify how the Temporal Archetype Bank is separated from the CiK evaluation data, particularly when they share the same data sources? A brief description of the time ranges and data splits would be helpful.

4 comments

Reply to W1: Methodological novelty

Thank you for this comment. The idea that context can improve forecasting is indeed long-standing and intuitive. Our research is motivated precisely by the substantial gap between this idea and the practical ability to realize it: effectively incorporating heterogeneous text into forecasting for series with different dynamics remains challenging. Across settings, what changes is not only the numerical distribution, but also how text conveys information and how that information should influence the forecasting process.

The semantic understanding of LLMs and the execution capabilities of agents offer a new approach to this problem. ECLIPSE makes this concrete by allowing context to determine which historical observations forecasting should rely on, which reference patterns are relevant, and how future trajectories should be adjusted. It translates these judgments into the selection, construction, and execution of programs while retaining the numerical modeling capabilities of TSFMs. We fully acknowledge the established foundations of the individual components. Our intended contribution is a method that coordinates these capabilities to address the alignment between contextual information and time-series computation, together with empirical evidence of its effectiveness.

Reply to W1 & Q1: Computational budget, attribution, and RCRPS

Thank you for this suggestion. We added budget-matched comparisons and RCRPS evaluations covering all 355 instances. To directly examine the contribution of orchestration, we provided an enhanced single agent with the same parsed inputs, category-specific guidance, retrieved information, tools, and executor as ECLIPSE, allowing either one planning call or up to eight planning–execution–checking rounds. We report the budget-matched comparison before audit and additionally include ECLIPSE’s audited result:

MethodTotal calls/taskRCRPS ↓DTR ↓
Moirai00.3160.278
Fixed Template≤30.2190.251
Single Agent40.2100.245
RAG + Post-processing40.1830.216
Enhanced single agent: one planning call50.1800.207
Enhanced single agent: up to eight rounds6.80.1890.214
ECLIPSE (without audit)50.1720.201
ECLIPSE (with audit; revised executor, original audit retained)6.820.1720.200

With matched inputs and call counts, ECLIPSE retains the lowest numerical errors, although its difference from the enhanced single-call agent is not statistically significant: the paired RCRPS difference is −0.008, with a 95% CI of [−0.035, +0.014]. The multi-round agent consumes more calls and approximately 27.8K tokens without improving average performance. Thus, additional reasoning budget does not automatically translate into gains; meanwhile, a single agent given ECLIPSE’s contextual processing capabilities can achieve comparable performance.

Under RCRPS, the overall ranking with the revised executor remains broadly consistent. ECLIPSE significantly outperforms Moirai, the fixed template, and the original Table 3 single agent. Its difference from RAG + Post-processing is not significant (−0.011, 95% CI [−0.025, +0.004]). We also report results with the original executor separately: ECLIPSE, RAG + Post-processing, and the enhanced single-call agent obtain RCRPS values of 0.184, 0.183, and 0.187, respectively. Finally, the current audit modifies point forecasts but not sampled trajectories, so its point-forecast improvements do not directly improve RCRPS. These additions distinguish the effects of computational budget, executor revisions, and evaluation metrics.

Reply to Q2: TemporalBench evaluation protocol and the improvement on Causal Chambers-HU

Thank you for this question. Yes, we used the same dataset version and evaluation protocol as the original TemporalBench paper. The corresponding baseline results in Table 2 are taken from that paper.

The substantial improvement on Causal Chambers-HU mainly stems from how ECLIPSE handles series scale and relative changes. Our framework focuses on handling the effects of contextual information on time series. Because many contextual effects are expressed as proportions or relative changes, ECLIPSE uses scale-normalized computations in its numerical analysis, which also benefit historical understanding. Specifically, the pressure readings are approximately 96,044 Pa, while fluctuations are only a few pascals; many samples exhibit an end-to-end relative change below 0.05%. Compared with models that directly interpret raw numbers, ECLIPSE better assesses the relative magnitude of these changes, reducing the tendency to mistake small fluctuations for pronounced trends.

HU follows a self-contained numerical analysis path without invoking scenario classification, retrieval, or intervention modules. The gains therefore reflect the benefits of executable numerical analysis without requiring additional context. The T1 solving process does not access reference answers; these are used only for final scoring.

Reply to W2: Generalization across contextual settings

Thank you for this comment. The seven scenario categories were indeed informed by CiK, but they characterize how context participates in forecasting—for example, historical events, future constraints, explicit rules, and indirect reasoning—rather than depending on a particular dataset. New contexts can fall within existing abstractions while requiring different operations, temporal scopes, and parameters. Generalization across settings therefore does not necessarily require introducing a new category. The framework also supports custom scenarios and tools; the seven categories are not a closed set of task templates.

We additionally evaluated ECLIPSE on TemporalBench, whose event-informed forecasting tasks span retail, electricity, experimental control, and clinical domains, providing different information conditions from CiK tasks with explicitly stated effects. For example, PSML provides qualitative event descriptions without specifying effect magnitudes; Causal Chambers provides changes in control variables without explicit response equations; and MIMIC provides medications and doses without directly specifying the target variable’s response. These settings require the system to determine appropriate computations from less complete contextual information. They examine whether the same planning abstractions can accommodate different forms of information and domain mechanisms, rather than whether the system can match a particular CiK task template.

Reply to W2 & Q3: Numerical stability and executor safeguards

Thank you for this suggestion. The executor includes local formula-convergence checks and a configurable magnitude guardrail: when an intervention’s mean relative change exceeds the threshold, the adjustment is reverted to the pre-intervention forecast. The reported runs enabled the local convergence checks but not the general magnitude guardrail.

It is important to distinguish legitimate large changes from amplification caused by incorrect execution. For example, in the “fivefold withdrawals” task, the full program applies a localized fivefold adjustment, whereas the ablated program mistakenly compounds the multiplier daily, producing DTR of approximately 10⁷. Enabling the guardrail reverts the erroneous recurrence while retaining the correct adjustment.

The reference DTR for full ECLIPSE is 0.190. We replayed the saved ablation programs under different thresholds while keeping the programs and forecast inputs fixed:

AblationUnboundedThreshold 0.8Threshold 2Threshold 5
−Contextual Priors47,853.40.2380.2370.232
−Audit105.20.1910.1900.187
−Direct Context48.50.2060.1950.191

Simple bounds prevent these numerical divergences, and all six paired effects in Table 4 remain statistically significant at the three listed thresholds. At thresholds 2 and 5, ECLIPSE’s improvements over the TSFM are 0.0171 and 0.0184, respectively, compared with the original 0.0170.

However, this protection can also discard grounded interventions that the system can already execute. Exceeding the threshold causes the entire adjustment to be reverted; consequently, legitimate contextual corrections may be discarded when future events or physical relationships require large responses. For example, in PressureFromSpeed, rotational speed increases from 661 to 1590. The 0.8 threshold reverts the physically supported adjustment, increasing DTR from 0.258 to 0.882. SVAR equations and irradiance inference based on zenith angle and cloud conditions can likewise require large changes. Among physically or equation-driven tasks, the 0.8 threshold reverts 21 beneficial interventions, causing a cumulative DTR increase of 4.19; threshold 5 still reverts four, with a cumulative increase of 0.55. Preserving aggregate gains therefore does not mean preserving every attainable benefit from context: some are sacrificed by the threshold rule.

This also motivates the audit module: errors may involve direction, temporal scope, rule interpretation, or contextual grounding, and cannot be judged solely by magnitude. Some incorrect adjustments are small, while some correct operations must produce large changes. The audit checks consistency with context and constraints, complementing deterministic numerical safeguards. Simple bounds provide a numerical backstop but cannot replace checks of an intervention’s grounding and semantics.

Reply to Q4: Separation of the Temporal Archetype Bank from CiK

Thank you for this question. CiK contains 355 evaluation instances, allowing us to inspect each instance’s data source and historical and forecast windows. We have checked this information and separated the Temporal Archetype Bank through non-overlapping time ranges, different stations, and data splits. Sharing an upstream data source does not imply using the same evaluation windows.

For example, the solar irradiance data exclude 2022—the year used by CiK—and come from different stations. The Montreal fire and rescue data use only 2017 and 2020, neither of which overlaps the historical or forecast windows of the corresponding CiK instances. Other sources use their native training splits or training prefixes. Verification combines year, station, and split metadata with additional window-level numerical comparisons for series at matching frequencies. We will add a concise description of the time ranges and data splits to the revised manuscript.

Dear Reviewer 971o,

Thank you for your thoughtful review. We hope our new budget-matched comparisons and RCRPS evaluations, numerical safeguard analysis, and clarifications on the TemporalBench protocol and retrieval-bank separation address your concerns. If so, we would appreciate your consideration of a higher score. We are happy to answer any further questions.

Best regards,

The Authors

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

The paper presents an agent framework ECLIPSE, which utilized contextual information in a novel way, such as translating into executable operations, for time-series forecasting and question answering. It will involve scenario interpretation, temporal examples retrieval, future constraints enforcement and auditing and so on. They have done evaluation on CiK and TemporalBench. Such framework also supports temporal QA through structured evidence extraction. The overall presentation is understandable, though the result section does a bit more work for me to follow. The main claims are so missed in the between.

Critical strengths
  1. This is a practical and useful problem that presents an interesting solution in this paper for incorporating context.

  2. The paper has reported empirical analysis that covers both performance gain, but also the failure mode. It provides explicit information for distinction between different sources and modules.

Critical weaknesses
  1. The evaluation part is not convincing. For example, at least, the experiments do not fully separate the benefit of the proposed orchestration from the benefit of additional computation. While several comparisons share the same tools and forecasting backbone, ECLIPSE uses substantially more language-model calls and tokens. This makes the specific contribution of the agent architecture less certain.

  2. I wonder if the conclusion is strongly supported by the evidence. The generality and reliability of the contextual reasoning remain only partially established. The scenario categories are informed by the main benchmark, while the audit experiments focus substantially on selected or injected errors. The evidence therefore leaves questions about unfamiliar contexts and the frequency with which auditing unnecessarily changes valid forecasts.

Questions for authors
  1. The experiment protocol could not distinguish well the design of ECLIPSE. For example, how much of the improvement comes from the proposed agent design, and how much comes from additional computation? Specifically, Table 3 compares methods with a shared toolchain, but Table 19 reports substantially different call and token costs. Is there evidence that the advantage persists when a single-agent baseline receives a comparable budget for reasoning and revision? How stable are these differences across repeated planner runs?

  2. I wonder how well does the framework handle contextual situations beyond those used to develop its scenario categories? Specifically, which TemporalBench settings require contextual mechanisms not already represented by the CiK-informed categories? Were the scenario definitions, retrieval procedures, and intervention rules fixed before evaluating those settings?

  3. How reliably does the auditing mechanism distinguish harmful interventions from valid ones? Specifically, alongside the injected-error results in Table 18, how often does the auditor reject or revise an independently verified correct trace? Please also clarify whether the auditing model and configuration in this experiment match those used for the main forecasting results.

3 comments

Reply to W1 & Q1: Architecture, inference budget, and stability

Thank you for this comment. To further distinguish the contribution of the design from additional computation, we added stronger single-agent controls using GPT-5.4-mini on all 355 CiK instances. These controls share ECLIPSE’s parsed inputs, scenario categories and skill guidance, retrieved references, full operator set, and executor, with either one round or up to eight rounds of execution feedback and revision.

MethodRCRPS ↓DTR ↓
Single agent, same guidance, one round0.1800.207
Single agent, same guidance, multiple rounds0.1890.214
ECLIPSE planning and execution0.1720.201

Results use the official CiK weights and evaluate pre-audit outputs under the same revised executor; original-executor results are also reported. This control already shares ECLIPSE’s core contextual processing capabilities and examines the remaining planning increment. ECLIPSE remains numerically best, while additional self-revision did not improve average final performance.

The multi-round agent averaged 2.82 calls and approximately 27.8K tokens, compared with 7.8K tokens for the single-round agent. Round-wise diagnostics also show that most improvement occurred in the first round:

StageDTR, capped at 1 ↓
Backbone forecast0.222
Round 10.187
Rounds 2–40.191–0.192
Final output0.190

Furthermore, Section 4.2/Table 1 and Appendix C.4/Table 19 already report ECLIPSE’s performance–cost advantage over TimeClaw. With the same DeepSeek backbone:

MethodRCRPS ↓Calls/task ↓Tokens/task ↓
TimeClaw0.18810.9109.5K
ECLIPSE0.1747.221.8K

Costs were measured on a 25-task subset. ECLIPSE achieves lower error with approximately 34% fewer calls and 80% fewer tokens. This demonstrates the complete framework’s practical efficiency, while the new shared-capability controls further examine planning and iteration.

Regarding stability, each single-agent condition was run three times. On a fixed 100-instance subset, the between-run DTR standard deviations were 0.003 and 0.008, respectively. ECLIPSE uses one existing run as its reference. We explicitly state this scope and distinguish cross-instance confidence intervals from the as-yet unmeasured between-run stability of ECLIPSE itself.

Reply to W2 (generality) & Q2: Generalization across contexts through shared planning abstractions

Thank you for this question. The seven scenario categories were indeed informed by CiK, but they classify how context participates in forecasting, rather than CiK’s datasets or specific task types. We consider historical events, future constraints, explicit rules, and indirect reasoning to provide a reasonable and broad characterization of the main roles of context. We therefore see no need to introduce a separate taxonomy solely for TemporalBench. The framework also supports custom scenarios and tools, so the seven categories do not constitute a closed boundary on its capabilities.

Accordingly, TemporalBench’s additional challenges do not necessarily require categories beyond the seven. Instead, the same contextual category can require different computational mechanisms under different information conditions. For example, Causal Chambers provides changes in control variables without the explicit relationships supplied in the corresponding CiK tasks; PSML provides qualitative event descriptions without specifying effect magnitudes; and MIMIC provides medications and doses without directly specifying the target variable’s response. These settings require the system to determine the appropriate computation from less complete contextual information, rather than simply applying an existing task template.

The scenario definitions, retrieval procedures, and intervention rules were fixed before evaluation on TemporalBench. Within this fixed configuration, whether to trigger retrieval, which reference information to retrieve, and which operations, temporal scopes, and parameters to use are determined by each instance’s context rather than dataset identity. The instantiated computation therefore varies across examples while using the same planning abstractions. This is the cross-scenario capability we aim to examine: whether these abstractions can translate different forms of context into executable computations appropriate to the task.

Reply to W2 (reliability) & Q3: Audit false alarms and unnecessary revisions under the main experimental configuration

Thank you for this suggestion. We supplemented the injected-error experiments in Table 18 with an evaluation of independently human-validated traces, specifically examining whether the auditor unnecessarily modifies acceptable forecasts.

Two annotators independently assessed 60 natural traces without seeing the audit decisions or revised predictions, achieving an initial agreement rate of 95%. After adjudication, 54 traces were considered acceptable. Of these, 47 entered the audit stage under the main pipeline’s rules, while seven did not trigger auditing. Among the 47 human-accepted traces, four were flagged as unreasonable (8.5%). Of these four, only two were actually modified (4.3% of all 47 traces). Of these two modifications, one increased forecasting error (2.1% of all 47 traces). These are nested outcomes, rather than three separate groups of errors.

Both revised predictions were subsequently reviewed by humans and remained acceptable; neither revision introduced an unacceptable defect. However, one revision increased DTR. We therefore report whether a revised prediction remains valid separately from whether the revision improves forecasting accuracy.

Regarding the auditing model and configuration, Table 18 uses the same auditing procedure, model, and runtime configuration as the corresponding main forecasting experiments. Since the original DeepSeek model version has been discontinued, we additionally conducted injected-error experiments using GPT-5.4-mini. The natural-trace evaluation above directly uses the audit decisions and final predictions stored in the original GPT-5.4-mini main experiment, and therefore measures actual false alarms, modifications, and accuracy changes under the main experimental configuration.

Dear Reviewer pbeE,

Thank you for your thoughtful review. We hope our new single-agent controls, clarification of generalization beyond CiK, and evaluation of audit false alarms on human-validated traces address your concerns. If so, we would appreciate your consideration of a higher score. We are happy to answer any further questions.

Best regards,

The Authors

Reviewer_3DGwOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The submission addresses forecasting and question answering on time series accompanied by a textual context. That context carries information absent from the numerical history. The proposed system, ECLIPSE, does not use an LLM to generate forecast values. Instead, it uses an LLM to select edits applied around a frozen, pretrained time-series forecasting model, Moirai-2.0-R-small:

  • History reconditioning (before forecasting). The observed history is modified, for instance, by repairing a corrupted window. If the history is shorter than one cycle, a typical cycle retrieved from an offline library is inserted.
  • Forecasting. Moirai produces the forecast from the modified history.
  • Future scenario enforcement (after forecasting). The forecast is modified, for instance, by scaling a stated window or imposing a stated bound.

The edits are drawn from a fixed library. The LLM outputs the operation names and their arguments, and deterministic code executes them.

Critical strengths
  • The LLM determines which operations to apply and with which parameters. All numerical values are produced by the forecasting model or by deterministic operations.
  • In direct prompting, the same LLM receives the history and context and writes the forecast values itself. ECLIPSE improves on it with each LLM backbone tested (Table 11).
Critical weaknesses
  • Correcting a history for known past events and adjusting a forecast for known future events are established practices see Box & Tiao, 1975 or Fildes et al., 2009. LLM tool use, retrieval for time-series forecasting, and LLM-based verification with revision are each standard. The contribution consists in combining these elements with a fixed operation library.
  • The gap in performance may reflect instructions and compute rather than the pipeline’s structure. Indeed eclipse receives task specific guidance that the alternatives may not, and uses 7.2 LLM calls per task against one in other models tested, see Table 3 and Table 19.
  • Moirai scores very differently on the same 355 instances in Tables 11 and 3 (intemporal 0.291 vs 0.566, covariate 0.284 vs 0.402, causal 0.521 vs 0.907), so the reported improvements cannot be relied upon until this is explained
  • The seven scenario categories were designed around CiK’s task types (Section 3.3.1), and CiK contexts state the effect explicitly
Questions for authors

1.Did the alternatives in Table 3 receive the same category-specific guidance as ECLIPSE? If not, does ECLIPSE retain its advantage when they do, at a matched number of LLM calls? 2. What differs between the Moirai-only evaluations in Tables 11 and 3 (context length, point-forecast definition, sampling, or instance set)? 3. How does ECLIPSE perform on tasks where the context implies an effect without stating its size or timing, for example news-conditioned financial forecasting?

5 comments

Reply to W1: From contextual understanding to executable forecasting programs

Using context to improve forecasting is a longstanding and intuitive idea. Our motivation is the gap between this idea and its practical realization. Events, background knowledge, and domain constraints clearly matter, yet making them meaningfully influence forecasts remains challenging when textual descriptions are heterogeneous and time series follow different dynamics. Across scenarios, what changes is not just the numerical distribution, but also how information is expressed and how it should affect forecasting. Giving a TSFM access to context does not, by itself, resolve these challenges.

LLM semantic understanding and agent execution offer an opportunity to address this gap. ECLIPSE makes this opportunity concrete by allowing textual meaning to govern how time-series computation proceeds. It translates scenario understanding into executable operations on history, reference patterns, and future trajectories, while retaining the numerical modeling capabilities of specialized forecasting models. The central design choice is that context can determine the historical state on which a forecaster operates and the conditions its output should satisfy, beyond serving as an additional predictor input.

We acknowledge the established foundations in intervention analysis, forecast adjustment, retrieval, and verification. ECLIPSE’s contribution is a computational interface connecting these capabilities for contextual time-series tasks, enabling the system to select, construct, and execute programs appropriate to each scenario. An idea can be simple to state without its implementation following directly from existing components. We contribute a method for realizing this idea across diverse contextual tasks and experimental evidence for evaluating it.

Reply to W2 & Q1: Category-specific guidance and inference budget

Table 3 compares contextual processing under different module configurations. ECLIPSE’s category-specific guidance (skills) was co-designed with the complete agent’s planning and execution workflow. Its recommendations are implemented through history processing, reference-pattern utilization, and future interventions, making this guidance part of the methodological design.

In the original comparison, including RAG+Post, we retained as much skill content as was compatible with the available modules, but did not provide the complete category-specific guidance used by ECLIPSE. Removing modules can also leave some recommendations without corresponding execution steps; copying the full guidance therefore does not guarantee that it functions as it does in the complete workflow. We distinguish between providing identical guidance content and retaining the modules needed to implement it. The additional experiments directly control the former to examine how much of the performance difference guidance alone can explain.

Specifically, we retained SA’s and RAG+Post’s original methods, prompt templates, and action spaces, supplied the same category-specific guidance, and tested three planning settings: initial planning only, initial planning plus up to two revisions, and initial planning plus up to five revisions. Each revision opportunity provides execution feedback, after which the model may revise or submit its solution. Experiments used GPT-5.4-mini on all 355 instances, with three runs per condition for each baseline, aggregated using the official CiK task weights:

MethodPlanningCallsRCRPSDTRMedian DTR
SA: original settinginitial planning only40.2100.2450.143
SA + guidanceinitial planning only50.2140.2490.142
SA + guidance+ up to two revisions7.00.2120.2470.142
SA + guidance+ up to five revisions7.00.2130.2490.142
RAG+Post: original settinginitial planning only40.1830.2160.137
RAG+Post + guidanceinitial planning only50.1800.2120.137
RAG+Post + guidance+ up to two revisions6.20.1930.2270.140
RAG+Post + guidance+ up to five revisions7.50.2020.2370.140
ECLIPSE (with audit, revised executor)full workflow6.820.1720.2000.104

Actual total calls include preprocessing and other stages. Revision counts are upper limits; no further revision calls are made after submission. Adding guidance leaves SA’s performance largely unchanged and modestly improves RAG+Post; additional execution feedback and revision opportunities do not further improve average performance. At approximately seven actual total calls, ECLIPSE (6.82 calls) achieves lower RCRPS, DTR, and median DTR than SA (7.0 calls) and RAG+Post (7.5 calls). These results support that ECLIPSE’s advantage cannot be explained simply by additional category-specific prompting or more calls: supplying the same guidance to the existing baselines and expanding their revision budgets remains insufficient to reproduce the complete workflow’s performance.

Reply to W3 & Q2: Consistency between Tables 3 and 11

Thank you for highlighting this inconsistency. We checked five runs and confirmed identical instance-level Moirai DTR values for the same instances. The discrepancy arose from the evaluation sets and aggregation procedures, rather than changes in Moirai’s predictions.

Because of time constraints before submission, the original Table 3 experiment covered an analysis subset rather than the full benchmark. It reported 74 instances, but Table 5 incorrectly listed 355. Its Overall column used instance averaging, whereas its context-specific columns used CiK task weights. Table 11 used all 355 instances and CiK task weighting throughout. We apologize for failing to synchronize these settings and their descriptions during manuscript revisions.

After submission, we extended the comparison to all 355 CiK instances. Since the original DeepSeek model version was discontinued by the provider, we conducted these additional experiments using GPT-5.4-mini. The revised Table 3 consistently applies CiK task weighting to every column. Its Moirai baseline is now 0.278 overall, with context-specific scores matching Table 11 exactly. We have corrected the captions, Table 5, and the appendix explanation, including removing the incorrect attribution to baseline reruns. All corresponding changes are marked in blue.

The corrected comparison retains the aggregate finding: ECLIPSE achieves lower DTR (0.200) than Moirai (0.278), the fixed-template control (0.251), and retrieval-plus-post-processing (0.216). We have updated the context-specific discussion accordingly: ECLIPSE performs better on intemporal, future, and covariate information, ties Moirai on historical information, and performs comparably to retrieval-plus-post-processing on causal information.

Reply to W4: Scenario abstractions and reasoning beyond explicit contextual effects

The seven scenario categories were informed by contextual forecasting tasks, including CiK, but serve as planning abstractions that guide computation rather than reproducing CiK’s original information categories or task templates. They describe how context participates in forecasting—for example, through historical events, future constraints, or indirect analogies—while the specific operations, temporal scope, and parameters must still be determined for each instance.

Not all CiK tasks explicitly specify the effect. Section 3.2 of the original CiK paper distinguishes descriptive context from context that explicitly specifies an impact, while Appendix A.3 includes deductive and analogical reasoning tasks requiring inference beyond directly stated information. For example, solar forecasting requires inferring a future trajectory from the plant’s characteristics, climate similarities across locations, and limited history; the text does not directly provide the complete numerical effect. CiK therefore evaluates both the application of explicit information and forecasting based on background knowledge and indirect cues.

We also evaluated ECLIPSE on TemporalBench, covering contextual reasoning, event-informed prediction, and historical understanding across multiple datasets. These evaluations extend beyond CiK’s forecasting setting and provide complementary evidence of the framework’s ability to handle different contexts and task formats. Among these tasks, contextual reasoning and event-informed prediction are the most directly relevant to this concern.

Reply to Q3: Implicit and ambiguous news context

Thank you for this question. We would like to clarify the scope of this work: ECLIPSE aims to translate context into grounded, executable time-series computations through a lightweight framework. Our current focus is operational forecasting in settings such as retail, supply chains, and energy, where business events, operating mechanisms, and verifiable constraints can inform prediction. Reliably handling arbitrary contexts across all domains and noise levels is a longer-term goal, rather than a problem we claim to have solved here.

For ambiguous events without reliable evidence about magnitude or timing, our financial configuration adopts a more conservative intervention policy: it limits adjustments lacking quantitative support and preserves more of the numerical baseline. This reflects our design principle that context should influence forecasts when it provides a sound basis, while its impact should be constrained when evidence is insufficient.

To test this setting directly, we conducted two independent evaluations after freezing the configuration: BTC news-conditioned forecasting and EDT company-announcement forecasting, both with three-step horizons. All news was available before the forecast origin. For this asset-price setting, we used a last-observed-price baseline and included an otherwise identical no-news condition to distinguish the effects of baseline selection and news.

DatasetPersistenceMoiraiTimeClaw + newsECLIPSE w/o newsECLIPSE w/ newsNews − no news, 95% CI
BTC3.4674.1973.9203.4673.442−0.02 [−0.20, +0.18]
EDT3.6573.7053.8323.6473.751+0.10 [−0.03, +0.25]

The metric is three-step nMAE, expressed as a percentage of the last observed price; lower is better.

ECLIPSE maintains competitive forecasting performance in these tests. On BTC, its news-conditioned error is lower than Moirai’s and TimeClaw’s; on EDT, it is close to Moirai and numerically better than TimeClaw. Neither news-versus-no-news difference is statistically significant, so these experiments do not establish a reliable additional benefit from news. The limited average influence of news is consistent with the conservative policy of constraining adjustments under weak evidence.

Financial news also introduces an effect-calibration problem: identifying an event does not directly determine its price impact. The response depends on market expectations, information already reflected in prices, and the duration of the effect. The Temporal Archetype Bank provides a direction for extending our approach by connecting textual cues to retrievable numerical evidence. Future work could develop richer event–condition–response memories and explore domain adaptation, pretraining, or post-training to estimate effects in more open-ended contextual settings.

These tests clarify the behavior of our current lightweight approach and provide a concrete direction for extending it to broader contextual forecasting problems.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.