- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes state-conditioned control (SCC) for frozen small language model agents. Expert trajectories are compiled into procedural structures that support current-state guidance, action restrictions, and direct execution of supported steps. With environment-specific adapters, SCC improves over four fixed offline memory adaptations on ToolSandbox, ALFWorld, and WebShop. Experiments include soft/hard channel ablations, online resource accounting, and reuse with two additional small models.
Critical strengths
-
Useful improvements with frozen small models. SCC exceeds the strongest evaluated memory adaptation by 13.12 ToolSandbox score points, 61.94 ALFWorld success-rate points, and 3.80 WebShop success-rate points (Table 1). These results demonstrate practical value for the implemented harnesses without weight updates or online expert calls.
-
Informative channel ablations. The four configurations share an environment core, initial states, and native-action budgets. Hybrid improves over Soft by 10.04 points on ToolSandbox and 24.63 percentage points on ALFWorld, while the corresponding WebShop interval includes zero (Table 11). This provides useful evidence about where the combined hard channel adds value.
Critical weaknesses
-
The contribution of experience-derived decision allocation remains insufficiently isolated. SCC combines manually specified grammars, binding templates, and environment-specific controls (Table 7; Appendix C.4). These include ToolSandbox stopping guards, ALFWorld stage/object restrictions, and WebShop’s predefined option-before-purchase rule. The channel ablations switch these bundles together; they do not compare against generic guards alone or an experience-free scripted controller with comparable state information. Consequently, substantial gains could come from developers selecting control mechanisms suited to each environment, rather than specifically from the proposed use of expert experience to allocate decisions by state.
-
The evaluation does not establish generalization beyond development-exposed tasks. Appendix A.1 states that the evaluated suites informed benchmark/model selection and refinements to guidance and constraints. Excluding 95 ToolSandbox source-overlap IDs preserves a 12.78-point gain, but related templates and development exposure remain (Appendix A.2). Reusing the controller with different models on these same suites does not resolve this issue. Although the paper acknowledges this limitation, independent evaluation is particularly important for a method with substantial manual controller design; otherwise, the reported improvements may depend on adaptation to the evaluated suites.
Questions for authors
-
Attribution evidence: With generic stopping/repetition guards held fixed, how does SCC compare with an experience-free state controller using the same public information? This comparison is needed to distinguish the contribution of trajectory-derived dependencies from manual control design.
-
Generalization evidence: Can you provide results on tasks or procedure families held out throughout controller development, with rules frozen before evaluation? Such evidence would directly address the development exposure described in Appendix A.1.
-
Baseline clarification: Why omit AMD’s student-side subtask decomposition (Table 5), which can operate over a frozen memory store? Does restoring it, while charging its auxiliary inference cost, materially change the comparison?
-
Cost clarification: Appendix F.2 excludes experience collection and store construction. What are these one-time costs, and how many subsequent tasks are needed to amortize them relative to direct expert execution?
3 comments
Response to Reviewer AqVb - Part 2
Q3: Baseline clarification: Why omit AMD’s student-side subtask decomposition (Table 5), which can operate over a frozen memory store? Does restoring it, while charging its auxiliary inference cost, materially change the comparison?
Response: We added AMD's student-side decomposition experiment. SCC still achieves higher scores and lower token costs on TS and ALF, leaving the paper's experimental conclusions unchanged.
| Method | TS mean score | TS tokens/task (k) | ALF success rate (%) | ALF tokens/task (k) |
|---|---|---|---|---|
| AMD (original baseline) | 43.75 | 19.94 | 14.18 | 28.32 |
| AMD + decomposition | 42.59 | 14.98 | 18.66 | 26.58 |
| SCC | 56.54 | 4.59 | 82.09 | 14.53 |
Q4: Cost clarification: Appendix F.2 excludes experience collection and store construction. What are these one-time costs, and how many subsequent tasks are needed to amortize them relative to direct expert execution?
Response: The estimated one-time teacher API cost for TS is CNY 139.34, which can be amortized after approximately 177 tasks relative to direct teacher execution. The teacher is DeepSeek-v4-pro, and the estimate uses recorded usage and a common reference rate.
| Item | TS |
|---|---|
| Experience-collection API cost | CNY 139.34 |
| Mean API cost of direct teacher execution | CNY 0.7879/task |
| Tasks to amortize API cost | 177 |
The experience store is compiled by a deterministic program with no additional generative-model calls. TS adapter development took an estimated 64.7 minutes; this development effort is excluded from the teacher API cost above. SCC makes no teacher calls at runtime.
Thank you for the authors’ response. However, I still do not believe that the paper meets the standard for acceptance at this conference, and I will therefore retain my current score.
Response to Reviewer AqVb - Part 1
Thank you for acknowledging the practical effectiveness of SCC, as well as the analytical value of channel ablation across various control configurations. In response to your suggestions, we provide further details below regarding our experimental results and the corresponding statistical metrics, covering: (1) the contributions of empirical and manual control mechanisms; (2) task partitioning and the scope of generalization; (3) comparisons following the decomposition of the student-side AMD; and (4) one-time construction costs versus their amortization.
W1: The contribution of experience-derived decision allocation remains insufficiently isolated. SCC combines manually specified grammars, binding templates, and environment-specific controls (Table 7; Appendix C.4). These include ToolSandbox stopping guards, ALFWorld stage/object restrictions, and WebShop’s predefined option-before-purchase rule. The channel ablations switch these bundles together; they do not compare against generic guards alone or an experience-free scripted controller with comparable state information. Consequently, substantial gains could come from developers selecting control mechanisms suited to each environment, rather than specifically from the proposed use of expert experience to allocate decisions by state.
Q1: Attribution evidence: With generic stopping/repetition guards held fixed, how does SCC compare with an experience-free state controller using the same public information? This comparison is needed to distinguish the contribution of trajectory-derived dependencies from manual control design.
Response: With the public interfaces, state tracking, and generic guards retained, full SCC improves over the configuration without the graph-based procedure and its dependent mechanisms by 13.15 points on TS and 69.40 percentage points on ALF.
Stage guidance reads graph nodes, dependencies, and bindings, so it is removed together with the graph-based procedure and its dependent execution. TS retains tool-return fields and candidate arguments, while ALF retains native observations and internal tracking:
The two comparisons use the same 420 tasks and 134 games, respectively. The 95% CIs for the gains are [8.00, 18.63] and [61.19, 77.61]. They examine the combined contribution of the graph-based procedure and its guidance, constraints, and execution mechanisms.
W2: The evaluation does not establish generalization beyond development-exposed tasks. Appendix A.1 states that the evaluated suites informed benchmark/model selection and refinements to guidance and constraints. Excluding 95 ToolSandbox source-overlap IDs preserves a 12.78-point gain, but related templates and development exposure remain (Appendix A.2). Reusing the controller with different models on these same suites does not resolve this issue. Although the paper acknowledges this limitation, independent evaluation is particularly important for a method with substantial manual controller design; otherwise, the reported improvements may depend on adaptation to the evaluated suites.
Q2: Generalization evidence: Can you provide results on tasks or procedure families held out throughout controller development, with rules frozen before evaluation? Such evidence would directly address the development exposure described in Appendix A.1.
Response: We have tracked the development batches used for data collection alongside the IDs of the current evaluation tasks; any overlap is strictly limited to interface debugging. SCC reuses procedural experience and rules within the same environment, with objects and parameters dynamically rebound based on current observations. Consequently, the primary experimental results represent a large-scale generalization from the development set to the test set. The rules for selecting task IDs for experience collection are as follows: for TS and ALF, tasks are stratified by task family, and two batches of development tasks are selected using a fixed hash-based ordering; for WS, a fixed range of development goals is used. The current evaluation sets for TS exclude the specific task IDs used for experience collection, while for ALF and WS, the training sources and official evaluation sets are utilized. The results for tasks not included in the experience collection process are as follows: