acceptodds
Withdrawn from ICLR 2027

Anchor-and-Frontier: Scheduling Tool-Schema Exposure Along Agent Trajectories

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Tool-using language models repeatedly ingest executable schemas while acting, creating a trade-off between full-catalog exposure and compact-context efficiency. In multi-step trajectories, however, an omitted tool can alter subsequent calls, observations, and decisions, making schema exposure a trajectory-level scheduling problem rather than only a per-turn retrieval problem. We introduce Anchor-and-Frontier (A&F), a deterministic controller that restores broad schema context at observable user-turn and execution-error checkpoints and otherwise exposes a cardinality-capped, state-conditioned frontier, without additional generative-model calls. On the BFCL V4 primary split, an always-compact controller reduces model-token use by 19.66% but loses 21.57 task-success percentage points relative to full-schema exposure, motivating selective broad-context restoration. On a predeclared 394-example condition holdout, Qwen3.5-27B with A&F achieves 259/394 successes versus 258/394 for full schema while reducing model tokens by 15.85%; a later same-condition Qwen3.5-122B analysis shows a similar 15.55% reduction. Across additional analyses and transfer experiments, token savings are more consistent than success retention, indicating that the efficiency–reliability trade-off is setting-dependent.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
LLM agents, tool use, function calling, tool retrieval, tool-schema exposure, trajectory scheduling, context efficiency, agent evaluation

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

The paper proposes Anchor-and-Frontier, an inference-time controller that dynamically schedules tool-schema exposure based on observable interaction events. It restores full tool context at key checkpoints, such as new user turns and tool execution errors, while using a compact, state-conditioned tool set at other steps to reduce token consumption while preserving end-to-end agent task success.

Critical strengths
  1. The paper highlights a limitation of single-step tool retrieval, which fails to fully account for the cascading effects of tool omissions on subsequent execution and decisions. The proposed A&F dynamically adjusts tool-schema exposure based on observable interaction events, offering a practical approach to improving context efficiency in multi-step agent trajectories.

  2. The paper conducts end-to-end evaluations on BFCL V4. On a predeclared set of 394 test cases, A&F achieves task success comparable to full-schema exposure with Qwen3.5-27B while reducing model-token consumption by 15.85%, providing empirical evidence of its efficiency benefits.

Critical weaknesses
  1. Insufficient isolation of Anchor scheduling. The experiments do not disentangle the effects of Anchor scheduling, the state-conditioned router, and the next-tool prior. Always-Compact and A&F are evaluated on different splits with different configurations, and comparisons with simpler strategies, such as periodic or random restoration, under comparable token budgets are missing.

  2. Limited generalization due to scenario overlap. Some test scenarios overlap with those used during development, while the next-tool prior is learned from successful development trajectories. Thus, the reported performance may partly rely on previously seen scenario patterns rather than generalization to unseen tasks.

  3. Limited novelty and unclear advantages over prior work. DTDR, MemTool, and ITR have already explored dynamic tool retrieval and context management. A&F mainly introduces event-triggered schema restoration, but its advantages remain unclear without direct comparisons with these approaches.

Questions for authors
  1. Can the authors isolate the contribution of Anchor scheduling through comparisons with periodic or random restoration under the same router and comparable token budgets?

  2. How well does A&F generalize to entirely unseen scenarios, particularly without relying on the demonstration-derived next-tool prior?

  3. What advantages does A&F offer over existing approaches such as DTDR, MemTool, and ITR, and can these be demonstrated through direct comparisons?

Reviewer_v8E3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a deterministic controller, namely the Anchor-and-Frontier (A&F) strategy. When invoking tools, this strategy reduces model token consumption while trying to limit the drop in success rate relative to full-schema exposure. It mainly addresses the efficiency–reliability trade-off: whether to fully expose the entire tool catalog or maintain a compact context for better efficiency.

Critical strengths
  1. The efficiency–reliability trade-off is important in realistic scenarios
  2. The proposed method is clear, the main idea: user turns and execution errors trigger broad exposure while a compact frontier is used at other times, is intuitive and easy to reproduce.
Critical weaknesses
  1. The experiments are mainly conducted on the BFCL benchmark. The results only demonstrate that the proposed strategy works under a fixed model, fixed executor, fixed organization of tool schemas and a limited set of tasks. Its generalizability remains insufficiently validated, for example on longer trajectories and across more model families.
  2. The baselines are not strong enough. The main baselines adopted in the paper are full schema and several compact routers, while lacking an adaptive-scheduling baseline. Additionally, TF-IDF, BGE-M3 and the basic compact router do not represent strong multi-turn schema scheduling methods
  3. Several comparisons are unfair. For ToolSandbox, the A&F strategy directly transfers the schedule learned from BFCL instead of being selected independently on ToolSandbox; Static-16 and A&F have different actual exposure budgets; In Table 13, A&F consumes substantially more tokens than Static-16, but the success rate is only improved under partial settings;
  4. The experiments are only conducted on Qwen3.5-27B and 122B-A10B. It remains unclear whether the proposed method generalizes to other model series such as Claude, Gemini, Kimi, GLM and DeepSeek.
Questions for authors

Please refer to the weaknesses. We encourage the authors to improve the completeness of the paper.

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

This paper studies when an agent should see the full tool catalog and when a smaller set is enough. Anchor-and-Frontier shows all available tool schemas after a new user turn or an execution error. Between these checkpoints, it uses a compact router that combines semantic matching, tool dependencies, and a next-tool prior learned from demonstrations. On a BFCL condition holdout, the method reduces model tokens by about 16% while meeting a prespecified five-point success-rate tolerance. Additional experiments examine scenario overlap, prior dependence, and transfer to ToolSandbox, where success retention is less conclusive.

Critical strengths

The method is easy to understand and follow. On the main BFCL condition holdout, A&F reduces total model tokens by about 16%, with similar aggregate task success. The authors report paired confidence intervals and distinguish condition generalization from generalization to new scenarios. They also report task-level regressions and the weaker ToolSandbox results.

Critical weaknesses

Major

  1. The methodological contribution is limited. The paper studies a specific part of dynamic tool selection: when to restore the full catalog. Its main addition is a rule that restores tools after user messages and execution errors. Dynamic tool exposure and fallback already appear in prior work discussed in Section 2. The paper does not clearly establish what broader insight this particular rule adds. The scope is narrow, and the evidence for a meaningful advance over existing approaches is limited.

  2. The main experiments do not establish the value of the proposed checkpoints. The headline comparison is between the complete A&F system and full tool exposure. The always-compact results in Table 5 use a different split and controller, so they do not isolate the effect of the schedule. A comparison using the same final router and prior is missing. There is also no comparison with periodic restoration or a wider frontier at comparable token usage. It therefore remains unclear whether user/error checkpoints are particularly effective, or whether other ways of allocating the same exposure budget would work equally well.

  3. The positive evidence is concentrated in one BFCL setting. The main BFCL holdout changes interaction conditions but overlaps with development scenario identities. This is consequential because the router uses a prior derived from successful development traces. On the scenario-disjoint 27B subset, success drops from 79/120 to 72/120. Neither this subset nor the corresponding 122B subset establishes the stated non-inferiority margin. ToolSandbox does not provide clear confirmation either: success drops from 136/243 to 127/243, and the confidence interval for end-to-end token savings includes zero. The synthetic study and four-task catalog pilot offer limited additional evidence. The demonstrated benefit is therefore narrower than the range of settings motivating the work.

Minor

  1. The router is not fully specified in the paper. The exact fusion procedure, dependency extraction, and prior fallback are deferred to the pinned runtime in Appendix E. These details matter because the main result depends on the combined router and schedule. Their availability should be clarified.

  2. The practical efficiency benefit remains unmeasured. The paper measures model tokens but does not report routing latency or compute cost. Prompt caching could also affect the savings from reducing repeated schemas. The token result is useful, but its deployment benefit is still uncertain.

Questions for authors
  1. Do you have results comparing A&F with always-compact exposure using the same final router and prior? If not, what evidence supports the choice of user/error checkpoints over simply allocating more tokens to tool exposure?

  2. Beyond the specific checkpoint rule, what is the main methodological or empirical contribution over existing dynamic tool-context methods such as MemTool and ITR? Which result best demonstrates this difference?

  3. What features does the next-tool prior use, particularly its categorical fallback? Can it access scenario identifiers or other scenario-specific information? This would help interpret the main condition-holdout result given the overlap with development scenarios.

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

Related papers

Discussion (0)

Sign in to comment.