acceptodds
Under review as a conference paper at ICLR 2027

Toward Green Multi-Agents: Understanding and Tackling Dynamic Residual Decay via Cross-Lifecycle Graph

Abstract

As multi-agent systems adapt their collaboration across tasks, they retain information from earlier rounds that can become outdated while continuing to influence decisions. Reusing such history can propagate invalid evidence; recomputing all affected work increases inference cost. Existing work on topology, persistent memory, and dependency updates does not jointly determine which persistent historical influence is associated with declining current-task utility and which computations the current task needs. We present the first systematic study of Dynamic Residual Decay (DRD), a cross-lifecycle phenomenon where residual influence persists across multi-turn tasks while its positive utility for the current task decays. To tackle DRD while preserving reusable work, we propose the cross-lifecycle graph (CLG), which traces invalidation forward and current demand backward, stops historical tracing at valid replacements, and re-executes the selected dependencies in order. Across four models and three multi-agent task sets, historical influence persists after positive utility declines. In cross-task recovery evaluations, CLG achieves a macro-averaged normalized utility of 0.9781 across task–model cells, compared with 0.9485 under Global Rerun, while reducing model calls by 39.8%, tokens by 29.9%, and elapsed time by 33.9%. Selectively retiring invalid influence and retaining useful history offers a route toward green multi-agent execution that is computationally cleaner and more resource-efficient.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multi-Agent Systems, Graph Representation, Dynamic Residual Decay, Green AI

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

This paper investigates Dynamic Residual Decay (DRD), a phenomenon in which historical information continues to influence multi-agent decisions even after its positive utility for the current task has declined. To address this issue, the authors propose the Cross-Lifecycle Graph (CLG), which combines forward invalidation propagation and backward demand tracing to identify dependencies requiring recomputation while preserving reusable historical information. The paper evaluates DRD across four models and three benchmarks, including TravelPlanner, ScienceWorld, and MARBLE. End-to-end recovery experiments show that CLG achieves a task-normalized utility of 0.9781, compared with 0.9485 for Global Rerun, while reducing model calls by 39.79%, token consumption by 29.95%, and execution time by 33.86%. Additional experiments examine the effects of task drift, dependency quality, and demand scope on recovery performance.

Critical strengths
  1. Meaningful problem formulation. The paper identifies the mismatch between historical influence and current-task utility in dynamic multi-agent systems, offering a useful perspective on information reuse.
  2. Promising efficiency improvements. CLG combines forward invalidation and backward demand tracing to avoid unnecessary recomputation, achieving substantial reductions in model calls, tokens, and execution time.
Critical weaknesses
  1. Limited methodological novelty. CLG mainly relies on established graph reachability and dependency-slicing techniques. Its distinction from conventional demand-aware incremental computation is insufficiently demonstrated.
  2. Limited robustness and generalizability. End-to-end experiments cover only two models, with inconsistent utility improvements. Recovery performance also deteriorates substantially under incomplete dependency records.
Questions for authors

Q1. What is the key algorithmic novelty of CLG compared with conventional demand-aware incremental computation? Q2. How does CLG ensure reliable recovery when dependency records are incomplete or inaccurate? Q3. Why is end-to-end evaluation limited to two models, and what explains the performance degradation on MARBLE–GLM? Q4. Are graph construction and dependency-tracking overhead included in the reported efficiency results, and can actual energy savings be quantified? Q5. How well do the constructed cross-lifecycle task sequences represent naturally evolving, long-running multi-agent workflows?

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

This paper studies Dynamic Residual Decay (DRD), in which artifacts left by retired agents keep shaping later decisions after their utility for the current task has declined, and proposes the Cross-Lifecycle Graph (CLG), which recomputes only the states lying downstream of an invalidated source and upstream of a demanded output. DRD is measured with four models on chained TravelPlanner, ScienceWorld and MARBLE tasks, comparing matched executions in which earlier history is readable or revoked, and CLG is compared with Global Rerun, two reimplemented baselines and a forward-only variant on two models, together with a synthetic-graph study of demand and edge perturbations.

Critical strengths

The measurement design is careful. Recording actual consumption edges on versioned LangGraph checkpoints, and comparing matched branches in which the same history is readable or revoked, gives a counterfactual handle on historical influence, and Appendix C.4 goes a step further by testing it with content replacement. This instrumentation is useful in its own right for studying stale memory in multi-agent systems.

The resource savings are consistent. CLG uses fewer calls than Global Rerun in all six cells (25.69% to 58.47%, Table 6) under a shared accounting window, and the per-cell numbers reproduce Table 2 exactly, so the cost side of the claim looks solid.

Critical weaknesses
  1. My main concern is that the paper does not yet show directly that residual history harms later tasks. The mismatch measure in Figure 3 and Table 4 counts neutral as well as harmful outcomes, and the influence indicator records reads, adoption and provenance rather than a causal effect: in Appendix C.4, 13 of the 20 classifiable consumers change output when the old content is replaced. The TravelPlanner loss rates in Section 4.2 (18.8% at zero drift, 61.5% at high drift) give how often readable history lowers the score but not by how much, and the per-task Delta U values are not shown. Since DRD is the first contribution, this weighs most on my rating.

  2. The recovery results leave a cell-specific quality loss unexplained. For MARBLE-GLM, CLG scores 0.8646 on the official F1 against 0.9514 for Global Rerun (Table 6), a paired difference of -0.087 [-0.163, -0.014] (Table 8), even though all recorded recovery-completeness checks pass. The six-cell normalized index (0.9781 against 0.9485) hides this regression, and only MARBLE-DeepSeek has an interval above zero, all from one run seed, so on current evidence CLG matches Global Rerun at lower cost rather than improving on it.

Questions for authors
  1. [Experiment] Behind the Section 4.2 loss rates, could you report the paired scores with and without history, the resulting Delta U values with counts and uncertainty, and the official constraints failed in representative cases? Figure 7(c) gives no exact values or uncertainty.

  2. [Experiment] Could you give the exact drift distance formulas and weights (Appendix C.1) and plot individual Delta U against numerical distance within each benchmark, with the probability and size of harm estimated with uncertainty? Four bins cannot show whether the relationship is gradual, thresholded or an artifact of the binning.

  3. [Experiment] For TravelPlanner, could you add traced examples and a breakdown of losses by changed task field or failed official check, to show which old artifacts cause which planning errors?

  4. [Experiment] For MARBLE-GLM, could you show paired task-level outcomes by drift and the retained and recomputed states in the failing cases, explain why reusing valid state lowered quality, and repeat the comparison over several seeds?

  5. [Clarification] What is the exact event-level rule for the influence indicator, how do the counts split across reads, adoption and derived paths, and how often does perturbing old content change the final decision or score? Separating harmful, neutral and beneficial outcomes would also help, as would explaining why MARBLE influence in Figure 7(b) appears much lower than in Table 4.

Reviewer_FVx5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies Dynamic Residual Decay (DRD): information from earlier agent work can continue to influence later tasks after its benefit has declined. It measures influence and task utility across constructed sequences based on TravelPlanner, ScienceWorld, and MARBLE. To limit stale information without recomputing everything, the proposed Cross-Lifecycle Graph (CLG) traces dependencies forward from invalidated artifacts and backward from current demands, then recomputes the intersection. The paper reports lower call, token, and elapsed-time totals than Global Rerun, with mixed task-utility results across the six model–task settings.

Critical strengths
  • Separating whether old information still influences an agent from whether it helps the current task is a useful framing. The paper reports both quantities rather than treating access or reuse as evidence of value.
  • CLG turns invalidation and current demand into a concrete recovery procedure, and the paper evaluates quality and resource use separately. It also states that its correctness depends on the recorded dependency graph and reports a cell where CLG performs worse, which helps keep the method’s limits visible.
Critical weaknesses
  • The paper’s task chains are constructed and do not show how often DRD occurs in unmodified agents. The quality evidence is also mixed: the normalized 0.9781 vs. 0.9485 hides TravelPlanner intervals that include zero, descriptive ScienceWorld results, and a significant CLG loss in MARBLE–GLM. RQ3 uses one model-run seed (42) across multiple task–drift units. Read/adoption logs remain a proxy for influence; the intervention check found effects for only 13/20 classifiable consumers, and 11 GPT–MARBLE labels were revised after positive-utility cases were reviewed. Thus, neither prevalence nor the size of the benefit is secure.
  • The technical novelty and practical scope of CLG remain unclear. Its central operation—marking descendants of an invalidated node and intersecting them with ancestors of demanded outputs—is close to established program slicing and demand-driven incremental recomputation. The paper describes the agent setting as its distinction, but does not make the algorithmic difference sufficiently clear. The method also assumes a known invalid source, known demand, and sufficiently complete consumption edges; Theorem 1 explicitly relies on recording all task-relevant parents. These assumptions are central to whether the approach works outside the constructed setup.
Questions for authors
  1. Problem relevance and positioning: Can you show a natural trace from an unmodified, long-running agent workflow where an old artifact continues to influence decisions after it stops helping? How common is this failure, and what does CLG add beyond the compaction, memory, and progress-file practices now used by frontier labs?
  2. Assessment-changing evaluation request: Could you report independent-seed, per-cell RQ3 results with paired uncertainty, including ScienceWorld, and a sensitivity analysis using one consistent influence-labeling rule? This would show whether the aggreg

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.