acceptodds
Withdrawn from ICLR 2027

SPMR: Continual Tool Retrieval Learning via Multi-Path Structured Prompt Memory Recovery

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Tool retrieval determines which capabilities a large language model (LLM) agent can actually exercise at any given moment. Generative tool retrieval internalizes tool knowledge into model parameters, turning retrieval from context enumeration into a learned mapping. As deployed tool inventories keep growing, however, a parameterized catalogue turns every tool registration into a model update, and extending a generative tool retriever becomes a continual learning problem. Existing continual learning methods largely attribute forgetting to knowledge loss and respond by freezing or otherwise protecting historical parameters. We argue that preserved parameters do not imply preserved retrieval performance, and decompose retrieval failure into three separable mechanisms: Storage, whether historical parameters survive; Access, whether the router can still reach them; and Competition, whether newly added labels displace historical targets even when their logits are unchanged. Motivated by this view, we propose a prompt-based generative tool retriever with a structured prompt memory of box-indexed soft prompts and a multi-path recovery memory mechanism that restores access to historical tools displaced by newly added tools.On ToolBench and ToolRet, our method surpasses most continual learning baselines. These results indicate that forgetting in continual retrieval should be measured along storage, access, and competition separately, rather than attributed to a single memory problem.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Tool Retrieval, Generative Retrieval, Continual Learning, Prompt-based Memory

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_maDjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces SPMR for continual tool retrieval, with the goal of preserving historical tool retrieval performance as new tools are introduced incrementally. The authors break forgetting down into Storage, Access, and Competition, and develop a framework that combines box-indexed structured prompt memory, geometric routing, and adaptive multi-path recovery.

Critical strengths

Meaningful problem: Continual tool retrieval is a relevant problem for evolving LLM agents, since newly introduced tools may interfere with the retrieval of existing ones.

Interesting perspective: The Storage–Access–Competition decomposition offers an intuitive perspective on retrieval forgetting that goes beyond conventional parameter preservation.

Useful ablations: Experiments on ToolBench and ToolRet show that box-based routing and multi-path aggregation improve retrieval performance relative to their simplified variants.

Critical weaknesses

Limited empirical competitiveness: SPMR is substantially outperformed by CODA-Prompt on both retrieval accuracy and forgetting metrics across both benchmarks. Additional results in Appendix B further show that SPMR often ranks behind several baselines on Recall@5 and NDCG@5, which makes its overall empirical advantages unconvincing.

Insufficient validation of the core motivation: Although the paper decomposes forgetting into Storage, Access, and Competition, it lacks controlled experiments that quantify each component's individual contribution to retrieval failures. In addition, the proposed multi-path recovery resembles conventional top-k routing and prediction ensembling, which makes its distinct methodological contribution unclear.

Incomplete comparison with recent CL methods: The evaluated baselines mainly cover prompt-based and traditional continual learning approaches, overlooking recent LoRA-based methods such as ProgLoRA and JANUS-LoRA. Discussing these methods and evaluating appropriately adapted parameter-efficient baselines would strengthen the experimental comparison.

Unclear practical advantages: Despite its substantially lower accuracy than CODA-Prompt, SPMR offers relatively modest persistent-memory savings while requiring multiple backbone forward passes during inference. The paper lacks sufficient latency and computational-cost analysis to justify this trade-off, as well as comparisons with non-parametric retrievers that can incorporate new tools through index updates.

Questions for authors

Performance–Memory trade-off: Table 1 shows that CODA-Prompt substantially outperforms SPMR while requiring only slightly more persistent memory. Could the authors compare the two methods under a matched memory budget to determine whether SPMR offers a meaningful trade-off?

Forgetting mechanism: Figure 3 provides examples of multi-path recovery, but how frequently do Storage, Access, and Competition failures occur across the evaluation sets? What proportion of recovered queries can be specifically attributed to correcting routing failures?

Comparison with recent CL methods: Could the authors clarify how SPMR differs from recent parameter-efficient continual learning methods such as ProgLoRA, and whether its structured prompt memory offers advantages over an appropriately adapted LoRA-based approach?

Reviewer_ZU5gOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies continual tool retrieval. A parametric retriever has to register new tools over a sequence of tasks while still ranking all tools seen so far, and the task identity is not given at test time. The authors argue that forgetting in this setting is not only a matter of overwritten parameters. They split it into three conditions: Storage (the historical representation is unchanged), Access (the inference procedure still reads the path that holds it) and Competition (the tool still ranks in the topk as the label set grows).

SPMR keeps a Llama-3-8B backbone frozen. Tool descriptions are clustered into domains and intents in MiniLM space. Each intent region is an axis-aligned box that owns a soft prompt. A router scores the query against the boxes, the top-p box prompts are each run through the backbone, and a shared, append-and-freeze tool classifier's logits are averaged across paths. A random-forest policy picks pp per query.

Experiments split ToolBench (about 11K base tools plus three tasks of roughly 640) and ToolRet (about 14K base tools plus four tasks of roughly 2.4K). SPMR is compared with seven continual-learning and continual-retrieval baselines over 5 seeds. The authors report that SPMR beats most baselines, has low forgetting and uses less persistent memory than the strongest prompt-based baselines. Ablations cover the path budget, the box keys and post-hoc tool editing.

Critical strengths
  1. The framing is useful and interesting for routed parametric memories. When storage is frozen by construction, the remaining forgetting has to come from routing or from label competition, and the paper says so explicitly. The path-budget ablation supports part of this: with storage fixed, moving from one to ten paths raises ToolBench AIP from about 47 to 60 and cuts forgetting from about 4.6 to 2.8. This is the main reason the work could be of interest.
  2. The evaluation reports results over 5 seeds with standard deviations, on two benchmarks and two metrics, and gives the persistent memory cost of each method. This makes the comparison easy to check, which is how the problems below became visible.
Critical weaknesses

[W1] The results do not show an advantage over existing methods, and the text/interpretation overstates them.

  • CODA-Prompt is better than SPMR on every accuracy and forgetting column of Table 1, on both benchmarks. On ToolBench it gets AIP around 80 against about 58, with forgetting about 1 against about 2, while using only around 7% more memory. On ToolRet SPMR is fifth of eight on AIP, behind L2P, InsCL and DSI++ as well. Its forgetting there is about the same as PGP's, which uses about a fifth of the memory. The abstract's "surpasses most continual learning baselines" and the statement in the paperthat SPMR "beats" CODA on stability (Section 4.2) are not supported by the paper's own table.
  • The low forgetting comes from a weak starting point. SPMR's base-task recall before any continual step is about 52 on ToolBench and 62 on ToolRet, far below CODA and L2P. The claim that it is "the only method that ends above its base score" follows from how AtA_t is defined. AtA_t averages over tasks, and the later tasks are small, so they score well above the base task (around 70–77 versus around 47 in Table 8). The rise is not evidence of stability.
  • The headline SPMR numbers change across tables. ToolBench Recall@5 AIP is about 58 in Table 1, 60 in Table 2 (adaptive) and 61 in Table 3, and forgetting is about 2.2, 3.0 and 3.1. The final average is about 62 in Table 4 and 66 in Table 8. The Table 1 forgetting is lower than any configuration in the ablation.
  • The adaptive path policy adds nothing measurable. At about 6 mean paths it scores within a fraction of a point of fixed p=6p=6 (Table 8), and in Table 2 it is worse than fixed p=8p=8 or 1010 on both AIP and forgetting. Most of the multi-path gain is already there at p=3p=3. Each path is also a full forward pass of an 8B model, so SPMR costs several times the inference compute of single-pass baselines, and no latency or FLOPs are reported. The "Params" column hides this cost.

[W2] The Storage–Access–Competition analysis is never measured, and the baselines needed to justify the design are missing.

  • The decomposition is the first listed contribution, yet no experiment attributes any method's forgetting to storage, access or competition. Two cheap diagnostics would do this. First, oracle routing to the gold box fixes access, so whatever forgetting remains is competition. Second, ranking only among old tools removes competition. Without them, "access should be recovered" is an interpretation of one ablation. Figure 3 does not help much, because it shows queries selected precisely because top-1 fails and the ensemble succeeds.
  • The decomposition is also tied to SPMR's architecture. Competition, defined as rank ≤k\le k, is simply the retrieval criterion, so Eq. 3 is close to a tautology. Storage, as defined, fails for any fine-tuned baseline even when it still retrieves well. Access does not hold up either: in Figure 3, paths from other boxes vote for the gold tool, so reading the box that "stores" yy is not necessary.
  • The box memory seems to do little of the storing. Appendix C reports L2 routing accuracy of about 27%, and with p=1p=1 the base-task recall is about 24% (Table 8). If prompts from mostly wrong boxes recover the gold tool through the shared classifier, the gain may just be test-time ensembling over a frozen backbone and an append-and-freeze classifier. A baseline with that classifier and a single shared prompt or no prompt, or with pp random prompts, is needed.
  • There is no non-parametric baseline. SPMR already embeds every tool description with MiniLM. A dense retriever over those descriptions adds a tool with an index write and cannot forget, and ToolRet itself benchmarks many such retrievers. The paper never shows that the parametric, classifier-based setting does better than this, so the problem SPMR solves may not need to arise. A joint-training upper bound and a sequentially fine-tuned lower bound are also missing, and the text does not explain how the vision continual-learning baselines were adapted or tuned.
Questions for authors
  1. Which SPMR configuration produced Table 1? Why do its AIP, forgetting and final average differ from Tables 2, 3, 4 and 8?
  2. How does a dense retriever over the same tool descriptions (MiniLM or a ToolRet-trained embedder) compare on Recall@5 and NDCG@5 at each task? What does joint training on all tasks give?
  3. With oracle routing, and separately with ranking restricted to old tools, how much of SPMR's forgetting remains? Can the same split be reported for CODA or L2P?
  4. What do a frozen backbone plus the append-and-freeze classifier, with no prompts or one shared prompt, and an ensemble of pp randomly chosen prompts achieve?
Reviewer_kFUUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SPMR investigates continual tool retrieval when new tools arrive sequentially, historical training examples are unavailable and inference must rank all registered tools without task identities. It distinguishes storage, access and label competition as possible sources of retrieval degradation. The method organizes description-derived hierarchical boxes into a growing memory of soft prompts, freezes historical entries and aggregates predictions from multiple independently evaluated prompt paths. A learned budget policy selects the number of paths for each query. Experiments on ToolBench and ToolRet report retrieval accuracy, forgetting, backward transfer and additional persistent state. Multi-path inference substantially improves the single-path variant, although the strongest evaluated baseline retains higher accuracy and lower forgetting.

Critical strengths
  1. The problem is relevant, and task identities are unavailable at inference.
  2. Multi-path inference substantially improves the single-path variant.
  3. Main results report multiple seeds and forgetting metrics.
Critical weaknesses
  1. The equivalence in Eq. (3) is unjustified: unchanged representations and selecting the owning box are unnecessary when every path scores all tools.
  2. CODA achieves substantially higher accuracy and lower forgetting on both benchmarks.
  3. Reduced additional state does not establish efficiency without latency measurements.
  4. Budget calibration, data construction and configuration mappings are incomplete; adaptive gains over the best fixed budget are small and lack uncertainty estimates.
Questions for authors
  1. Can the authors revise the formalization and isolate access failures from label competition?
  2. What is the complete calibration objective and permitted data access?
  3. Can they provide full splits, task-performance matrices, computation-matched baselines and end-to-end costs?

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

Related papers

Discussion (0)

Sign in to comment.