Pith. sign in

REVIEW 3 major objections 7 minor 12 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

Steering vectors control tool use in LLMs despite tools living only in context

2026-07-08 23:51 UTC pith:M24QYB3Q

load-bearing objection Solid empirical finding on steering non-parametric tool use, but the random-direction control gap is real and undercuts the causal specificity claim. the 3 major comments →

arxiv 2607.05790 v1 pith:M24QYB3Q submitted 2026-07-07 cs.AI

Controlling Tool Use with Heading-Specific Activation Steering

Yuqi Chen , Vincent Siu , Yang Liu , Dawn Song , Chenguang Wang This is my paper
classification cs.AI
keywords activation steeringtool usenon-parametric representationlinear representation hypothesiscausal controllabilitylanguage modelsinterpretabilityheading anchors
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper asks whether a language model's decision to call an external tool — a behavior that exists only in the prompt context at inference time, with no encoding in model weights — can be controlled by manipulating internal activations. The authors extract steering vectors from the hidden states at structured heading positions (e.g., the token where the model decides between ### Reasoning and ### Code), computing the mean difference between reasoning-heading states and tool-heading states. They show that adding this vector to the residual stream at a specific layer suppresses tool calls below baseline, while projecting activations to be orthogonal to the vector amplifies tool calls above baseline — a bidirectional causal effect replicated across five open-source models and three task domains. Suppression works best where parametric reasoning suffices (math problems where the model could solve without a calculator) and hurts most where tools are genuinely necessary (time-sensitive questions, underspecified user intent). The effect also transfers partially beyond the literal heading token: a vector extracted under one output schema still suppresses tool calls under a different schema where the original heading string never appears.

Core claim

The central finding is a dissociation between causal controllability and geometric regularity. The mean-difference steering vector suppresses tool use far more effectively than a supervised linear probe direction, even though the probe achieves near-perfect classification accuracy (AUROC 1.000 vs. 0.945 for the mean-difference vector). In other words, the direction that best decodes tool-use intent is not the direction that best controls it. Furthermore, tool-invocation hidden states show diffuse, bimodal alignment with the suppression vector rather than the consistent negative alignment that a clean linear encoding would predict, and different tool types (code execution, search, user-clarif

What carries the argument

Heading-anchored steering vectors: mean-difference vectors between hidden states at ### Reasoning positions and tool-heading positions (### Code, ### Search, ### AskUser), applied at a single layer via activation addition or orthogonalization.

Load-bearing premise

The steering vector is built by taking the mean difference between hidden states at reasoning-heading positions and tool-heading positions, and the paper assumes this particular construction captures a causally relevant direction for tool-use decisions. The paper itself shows this direction diverges from the most linearly decodable direction, and its geometric properties are diffuse and bimodal — so why this specific extraction method works better than alternatives remains an

What would settle it

If the steering vector merely suppressed the literal heading token string rather than the underlying tool-use decision, the causal claim would collapse. The paper addresses this with cross-format and heading-rename controls showing partial transfer, but the transfer is asymmetric and schema-sensitive, leaving open whether the effect is fully about tool-use intent or partly about format-specific token suppression.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Inference-time tool-use control without retraining: models could dynamically suppress or encourage tool calls based on task domain, reducing unnecessary API calls and latency in production agent systems.
  • The dissociation between probing accuracy and causal effectiveness challenges the assumption that the most decodable internal direction is the most intervention-relevant one, with implications for how interpretability research selects directions for behavioral control.
  • The bimodal, diffuse geometry of tool-use representations may be a general signature of context-injected (non-parametric) behaviors, potentially distinguishing them from weight-encoded behaviors like refusal or sentiment in mechanistic interpretability studies.
  • Cross-tool steering vector analysis showing low feature overlap suggests that different tool types recruit distinct internal circuitry, which could inform architecture design for multi-tool agents.
  • The consistent normalized steering depth (0.45–0.60) across model sizes suggests tool-use representations stabilize at a predictable relative position in the network, which could simplify deployment across model scales.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. This paper investigates whether tool-use decisions in LLMs — a behavior that is non-parametric by construction, since tools exist only in context at inference time — can be controlled via activation steering. The authors extract steering vectors from hidden states at heading-anchor positions (e.g., the '###' token preceding '### Code', '### Search', '### AskUser') by computing the mean difference between reasoning-heading states and tool-heading states (Eqs. 1–3). They show across five open-source models and three domains (Math, Time, Intention from the SMART benchmark) that activation addition suppresses tool use below baseline while orthogonalization amplifies it above baseline (Table 1, Figure 1), constituting bidirectional behavioral control. The paper further conducts geometric analysis (§6) showing that the steering vector does not correspond to clean linear structure: tool-step hidden states exhibit diffuse, bimodal alignment with the suppression vector (Figure 4), and different tool types recruit largely distinct feature dimensions (Table 2, Figure 5). Controls in §5 and Appendix A address whether the effect is tied to literal surface strings (cross-format transfer, heading rename), and Appendix A.3 compares the mean-difference direction against probe-based, whitened, random, and shuffled-label alternatives on Llama-3.1-70B.

Significance. The paper makes a genuine contribution by extending activation steering to a non-parametrically grounded behavior, which is conceptually novel relative to prior work on sentiment, refusal, and truthfulness. The bidirectional control claim is supported by the orthogonalization results producing the opposite effect from activation addition. The cross-format transfer and heading-rename controls (§5, Tables 3–4) are well-designed and address the most obvious confound (literal string suppression). The geometric analysis in §6, particularly the finding that the most linearly decodable direction (probe AUROC 1.000) is not the most causally effective one (Table 5), is a substantive and honest empirical contribution. The paper is appropriately transparent about the gap between causal effectiveness and geometric regularity, framing it as an open question rather than overclaiming.

major comments (3)
  1. §5, Table 1 and Appendix A.3, Table 5: The random-direction control is the most load-bearing gap for the causal specificity claim. Table 5 shows that on Llama-3.1-70B (Math/Code), a norm-matched random direction reduces ToolAvgUse from ~3.53 (baseline) to 1.680 — a 52% reduction — compared to the mean-difference vector's 0.235 (93% reduction). This demonstrates a substantial non-specific perturbation component: adding any vector of comparable norm at the '###' token disrupts tool invocation considerably. The extracted vector is clearly more effective than random, but the specificity gap is characterized for only one model and one domain. The headline claim of 'causal control' across five models and three domains (Table 1) has no random-direction or shuffled-label baseline for the other four models or for the Time/Intention domains. This is a correctness-risk concern: without knowing how
  2. §3.2, Eq. (5) and §5, Table 1: The orthogonalization results — which support the 'bidirectional' part of the central claim — have no random-direction control at all. If projecting out a random direction also increases tool use (e.g., by disrupting reasoning and causing fallback to tools), the bidirectional causal interpretation weakens. The paper's own §5 controls (cross-format transfer, heading rename) address whether the effect is tied to literal surface strings but do not address whether it is tied to the specific extracted direction versus any direction of similar norm. For the orthogonalization arm specifically, there is no control reported in any model or domain. This is load-bearing because the abstract and conclusion state bidirectional causal control as a primary finding.
  3. §4, Steering Vector Construction: The steering vectors are extracted from only 20 training examples per domain (60 total), and the structured trajectories used for extraction are generated by GPT-4o rather than by the target models themselves. This raises a question about whether the extracted direction generalizes because it captures a model-invariant tool-use representation or because the heading-anchor format creates a structural bottleneck that is largely model-agnostic. The paper hints at the latter in §6.1 ('intervening at this structural bottleneck may be sufficient for behavioral control even when the underlying concept has no consistent direction'), but this hypothesis is not directly tested. A brief discussion of why GPT-4o-generated trajectories yield effective vectors for Mistral/Llama models would strengthen the causal interpretation.
minor comments (7)
  1. Abstract: 'heading-anchors positions' should be 'heading-anchor positions' (extra 's').
  2. §3.1, Eq. (2): The sign convention is initially confusing. The text states v_k points 'from the ### Reasoning state to the corresponding target tool-heading state,' but adding v_k is described as suppressive (shifting away from tool-heading). This is correct but could be stated more explicitly — the vector points toward reasoning and away from tool, so adding it pushes toward reasoning.
  3. Table 1: The Intention domain reports 'Missing Details Recovery' and 'Summarized Intention' as fractions (e.g., '48.84 / 21.70'), but it is unclear what the two numbers represent without consulting the IN3 dataset reference. A footnote or caption clarification would help.
  4. §5, Steering Layer Selection: The elbow-selection method is described qualitatively ('select the elbow of the resulting tool-usage curve'). A more precise criterion (e.g., second derivative threshold) would improve reproducibility.
  5. Figure 2: The y-axis label 'Average tool calls per query' could note the domain used for the sweep (Math/Code is implied but not stated in the caption).
  6. §6.1: The statement 'these similarities do not describe the responses intervened on in Section 5, since steering shifts the trajectory of each interaction' is important but buried. Consider foregrounding this caveat earlier, as it limits the extent to which the geometric analysis can explain the causal results.
  7. References: The Llama 3 reference (Grattafiori et al., 2024) has an extremely long author list that spans multiple pages. Consider using 'et al.' abbreviation per journal conventions.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for a careful and constructive report. The referee correctly identifies that our random-direction and shuffled-label controls are currently limited to one model (Llama-3.1-70B) and one domain (Math/Code), and that the orthogonalization arm lacks any random-direction control. We agree these are load-bearing gaps for the causal specificity claim and will extend the controls. We also agree that the GPT-4o trajectory extraction warrants explicit discussion. Below we address each comment.

read point-by-point responses

  1. Referee: §5, Table 1 and Appendix A.3, Table 5: The random-direction control is characterized for only one model and one domain. The headline claim of causal control across five models and three domains has no random-direction or shuffled-label baseline for the other four models or for the Time/Intention domains.

    Authors: The referee is correct that the random-direction and shuffled-label controls in Table 5 are limited to Llama-3.1-70B/Math. We agree this is a load-bearing gap for the causal specificity claim and will extend the random-direction and shuffled-label baselines to all five models and all three domains in the revised manuscript. We note that the existing 70B/Math data already shows a meaningful specificity gap: the mean-difference vector reduces ToolAvgUse to 0.235 (93% reduction) versus 1.680 for the random direction (52% reduction). The referee is right that the 52% random-direction effect is non-trivial and indicates a substantial non-specific perturbation component. We will therefore revise the language in the abstract and conclusion to qualify the causal control claim: the extracted vector is significantly more effective than random, but the effect includes a non-specific component, and the degree of specificity varies across models and domains in ways the current manuscript does not fully characterize. We will present the extended controls in a new table and discuss the implications for the causal specificity interpretation. revision: yes

  2. Referee: §3.2, Eq. (5) and §5, Table 1: The orthogonalization results have no random-direction control at all. If projecting out a random direction also increases tool use, the bidirectional causal interpretation weakens.

    Authors: This is a fair and important point. We did not include a random-direction control for the orthogonalization arm in any model or domain. We agree that this is load-bearing for the bidirectional causal control claim stated in the abstract and conclusion. In the revision, we will add random-direction orthogonalization controls across all five models and three domains. If projecting out a random direction also increases tool use (e.g., by disrupting reasoning and causing fallback to tools), we will report this honestly and qualify the bidirectional interpretation accordingly. We acknowledge that without this control, the current bidirectional claim is not fully supported by the evidence presented. Depending on the results, we may need to revise the abstract to state that orthogonalization with respect to the extracted direction increases tool use, while noting that the specificity of this effect relative to random directions is not yet established. We cannot pre-commit to the outcome of experiments we have not yet run, so this remains a standing uncertainty. revision: yes

  3. Referee: §4, Steering Vector Construction: Steering vectors are extracted from only 20 training examples per domain, and the structured trajectories are generated by GPT-4o rather than by the target models themselves. This raises a question about whether the extracted direction generalizes because it captures a model-invariant tool-use representation or because the heading-anchor format creates a structural bottleneck. The paper hints at the latter in §6.1 but does not directly test it.

    Authors: The referee identifies a genuine confound in our extraction pipeline. We agree that the current design cannot distinguish between two explanations: (1) the extracted direction captures a model-invariant tool-use representation, or (2) the heading-anchor format creates a structural bottleneck that is largely model-agnostic, making almost any contrastive direction extracted at that position effective. Our discussion in §6.1 hints at the latter but does not test it directly. In the revision, we will add an explicit discussion of why GPT-4o-generated trajectories yield effective vectors for Mistral/Llama models, and we will frame the structural-bottleneck hypothesis as the more likely explanation given the evidence. We will also note that the cross-format transfer results (Table 3) partially address this: the asymmetry of markdown→json (63.9% suppression) versus json→markdown (15.5%) suggests the effect is not purely format-agnostic but is tied to the specific heading-anchor structure. However, we acknowledge that a direct test—e.g., extracting vectors from target-model-generated trajectories and comparing effectiveness—would strengthen the interpretation. We will add this as an explicit limitation and future direction rather than claiming we have resolved the confound. revision: partial

standing simulated objections not resolved
  • The outcome of the extended random-direction controls for the orthogonalization arm is not yet known. If random-direction orthogonalization also increases tool use substantially, the bidirectional causal control claim in the abstract may need to be significantly weakened, and we cannot determine the magnitude of this effect without running the experiments.

Circularity Check

0 steps flagged

No significant circularity: steering vectors are extracted from training data and evaluated on held-out test data from an external benchmark; the bidirectional control claim is supported by non-tautological opposing effects of addition vs. orthogonalization.

full rationale

The paper's central claim is that heading-anchored steering vectors exert bidirectional causal control over tool use. The derivation chain is: (1) extract steering vectors as mean differences between hidden states at '### Reasoning' and tool-heading anchor positions (Eqs. 1-3) from 60 training-split examples; (2) apply via activation addition (Eq. 4) and orthogonalization (Eq. 5) at inference time; (3) evaluate on held-out test data from the SMART benchmark (Qian et al., 2025), an external benchmark not co-authored by the present authors. The steering vectors are constructed from training data and evaluated on separate test data, so the 'prediction' (suppressed tool use on test queries) is not forced by the construction. The bidirectional claim rests on addition suppressing and orthogonalization amplifying tool use — these are mathematically distinct operations (adding a vector vs. projecting it out) and their opposing behavioral effects are not tautological consequences of the construction. The geometric analysis (§6) uses unsteered prompts and independently characterizes the extracted vectors, even finding that the mean-difference direction diverges from the most linearly decodable direction (Table 5: probe AUROC 1.000 vs. 0.945), which is an honest non-circular finding. Self-citations (Siu et al., 2025a/b/c) appear in related work and methodology for standard steering techniques (contrastive activation addition, orthogonalization), but these are citations to methodological precedents, not load-bearing premises whose validity determines the paper's central claim. The paper's results stand on the external SMART benchmark evaluation and the non-trivial observation that addition and orthogonalization produce opposing effects. The missing random-direction controls across all models (noted by the skeptic) is a correctness concern about effect specificity, not a circularity issue — the extracted direction is not defined in terms of the evaluation outcome.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The paper introduces no new entities (particles, forces, dimensions, etc.). The steering vectors are computed from existing model activations using a defined procedure. The free parameters (layer, coefficient, training set size) are standard calibration choices for activation steering methods. The axioms are domain assumptions about the applicability of mean-difference extraction and heading-anchor intervention to non-parametric tool-use behaviors.

free parameters (3)
  • steering layer l* = per model: 18, 16, 23, 19, 45
    Selected by sweeping all layers and picking the elbow of the tool-usage curve (Figure 2). This is a per-model calibration step.
  • steering coefficient α = 1
    Fixed at α=1 for all experiments (§3.2) without sensitivity analysis or justification for this value.
  • number of training examples for vector extraction = 20 per domain, 60 total
    Chosen in §4; no analysis of how steering vector quality scales with this number.
axioms (3)
  • domain assumption The mean difference between hidden states at '### Reasoning' and tool-heading anchor positions captures a causally relevant direction for tool-use decisions.
    Eqs. 1-3 define the steering vector as this mean difference. The paper provides no theoretical justification for why this extraction method should capture a causally effective direction, and the geometric analysis (§6) shows it does not correspond to clean linear structure.
  • domain assumption Intervening at the heading-anchor token (###) is sufficient to influence the model's tool-use decision.
    §3 states the intervention is applied 'locally at heading-formation steps, where the decoder is about to select the next section label.' This assumes the decision is localized to this token position.
  • domain assumption GPT-4o-generated structured trajectories are representative of the target models' internal representations.
    §4 states steering vectors are extracted from trajectories generated by GPT-4o, but applied to five different open-source models. This assumes the heading-anchor representations are transferable across models.

pith-pipeline@v1.1.0-glm · 21375 in / 2905 out tokens · 482840 ms · 2026-07-08T23:51:44.578787+00:00 · methodology

0 comments
read the original abstract

Tool-augmented large language models extend their capabilities beyond parametric knowledge through external tools, but tend to invoke them unnecessarily. We investigate whether tool-use decisions have any stable internal representation that can be extracted and manipulated, a question that is non-trivial given that tools exist entirely in context at inference time and have no direct encoding in model weights. We show that steering vectors extracted from heading-anchors positions exert bidirectional causal control over tool-invocation behavior across five open-source models and three domains, suppressing unnecessary tool use most effectively in domains where parametric reasoning suffices. However, geometric analysis reveals that this causal effectiveness does not correspond to clean linear structure: tool-invocation steps exhibit diffuse, bimodal alignment with the suppression vector rather than the consistent negative alignment a linear encoding account would predict, and different tool types recruit largely distinct internal signatures with low cross-tool feature overlap. We hypothesize these geometric properties are indicative of the non-parametric nature of tools, and distinguish tool-use steering vectors from those extracted for parametrically grounded concepts. The relationship between this geometric irregularity and the observed causal effectiveness remains an open question.

Figures

Figures reproduced from arXiv: 2607.05790 by Chenguang Wang, Dawn Song, Vincent Siu, Yang Liu, Yuqi Chen.

Figure 1
Figure 1. Figure 1: Average tool calls per query across five models under three conditions. Base Model Tool Prompt is the prompt-only baseline without hidden-state intervention. Activation Addition subtracts the steering vector from the residual stream, suppressing tool use below the baseline. Orthogonalization projects hidden states to be orthogonal to the steering vector, removing the suppressive component and amplifying to… view at source ↗
Figure 2
Figure 2. Figure 2: Average tool calls per query as a function of normalized steering layer depth across five models. Tool usage drops sharply as the steering layer moves into the middle of the network before plateauing, forming a clear elbow. Annotated points mark the selected elbow layer for each model; despite differing in absolute depth, all five fall within normalized depths 0.45–0.60, suggesting that tool-use representa… view at source ↗
Figure 3
Figure 3. Figure 3: Average tool calls per query under orthogonalization and the base model prompt across five models and three domains. Orthogonalization projects hidden states to be orthogonal to the suppression vector at tool-heading positions, removing its suppressive component. Tool usage increases above the base model level in all cases, suggesting that the suppression vector is active in the unsteered baseline and that… view at source ↗
Figure 4
Figure 4. Figure 4: Cosine similarity between the steering vector and hidden states at reasoning steps (blue) and tool-invocation steps (red) on unsteered prompts, across five models. Reasoning steps concentrate sharply at zero, indicating near-orthogonality to the steering vector. Tool steps exhibit a diffuse bimodal distribution with mass on both the acute and oblique sides, inconsistent with the steering vector acting as a… view at source ↗
Figure 5
Figure 5. Figure 5: Pairwise cosine similarity between steering vectors extracted for code, search, and askuser tool types across five models. Code and search vectors are consistently more similar to each other than either is to the askuser vector, suggesting that clarification-seeking recruits distinct internal structure relative to external computation and retrieval [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages · 4 internal anchors

  1. [2]

    ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs

    URL https://aclanthology.org/2025. findings-acl.239/. Qin, Y ., Liang, S., Ye, Y ., Zhu, K., Yan, L., Lu, Y ., Lin, Y ., Cong, X., Tang, X., Qian, B., Zhao, S., Hong, L., Tian, R., Xie, R., Zhou, J., Gerstein, M., Li, D., Liu, Z., and Sun, M. Toolllm: Facilitating large language models to master 16000+ real-world apis, 2023. URL https://arxiv.org/abs/2307...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1007/s11704-024-40678-2 2025

  2. [3]

    LLM With Tools: A Survey

    URL https://aclanthology.org/2024. emnlp-industry.59/. Shen, Z. Llm with tools: A survey, 2024. URL https: //arxiv.org/abs/2409.18807. Siu, V ., Crispino, N., Park, D., Henry, N. W., Wang, Z., Liu, Y ., Song, D., and Wang, C. Steeringsafety: A systematic safety evaluation framework of representation steering in llms, 2025a. URL https://arxiv.org/abs/ 2509...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.18653/v1/2025.findings-acl 2024

  3. [4]

    RepIt: Steering Language Models with Concept-Specific Refusal Vectors

    URL https://aclanthology.org/2025. findings-acl.1310/. Siu, V ., Henry, N. W., Crispino, N., Liu, Y ., Song, D., and Wang, C. Repit: Representing isolated targets to steer lan- guage models, 2025c. URL https://arxiv.org/ abs/2509.13281. Turner, A. M., Thiergart, L., Leech, G., Udell, D., Vazquez, J. J., Mini, U., and MacDiarmid, M. Steering language model...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [5]

    SVD-LLM V2: Optimizing Singular Value Truncation for Large Language Model Compression

    Association for Computational Linguistics. ISBN 979-8-89176-189-6. doi: 10.18653/v1/2025.naacl-long. 11 Controlling Tool Use with Heading-Specific Activation Steering

    work page doi:10.18653/v1/2025.naacl-long 2025

  5. [6]

    naacl-long.331/

    URL https://aclanthology.org/2025. naacl-long.331/. Wollschl¨ager, T., Elstner, J., Geisler, S., Cohen-Addad, V ., G¨unnemann, S., and Gasteiger, J. The geometry of refusal in large language models: Concept cones and representational independence, 2025. URLhttps: //arxiv.org/abs/2502.17420. Wu, Z., Arora, A., Geiger, A., Wang, Z., Huang, J., Juraf- sky, D...

    work page doi:10.18653/v1/2025.emnlp-main 2025

  6. [7]

    ReAct: Synergizing Reasoning and Acting in Language Models

    URL https://aclanthology.org/2025. emnlp-main.898/. Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., and Cao, Y . React: Synergizing reasoning and act- ing in language models, 2023. URL https://arxiv. org/abs/2210.03629. Zou, A., Phan, L., Chen, S., Campbell, J., Guo, P., Ren, R., Pan, A., Yin, X., Mazeika, M., Dombrowski, A.-K., Goel, S., ...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [9]

    ‘‘‘ inside a ‘### Code‘ step when computation is needed

    Tool Usage: - Use code snippet ‘‘‘python ... ‘‘‘ inside a ‘### Code‘ step when computation is needed

    work page

  8. [11]

    ### Output Guidelines - Your answer must begin with ‘### Reasoning‘

    Goal-Oriented Resolution: - Conclude your reasoning process by achieving a clear, accurate, and succinct solution based on your independent analysis and any tool findings. ### Output Guidelines - Your answer must begin with ‘### Reasoning‘. - After that, every step must begin with one of these section headers on its own line: ‘### Reasoning‘, ‘### Code‘, ...

    work page

  9. [12]

    Use the tool only when necessary

    Reason Independently: - Leverage your own knowledge to analyze and solve reasoning steps whenever possible. Use the tool only when necessary

    work page

  10. [13]

    - Use a ‘### AskUser‘ step whose content is the exact question that should be asked to the user when user-specific clarification is needed

    Tool Usage: - Use a ‘### Search‘ step whose content is a concise web search query when external information is needed. - Use a ‘### AskUser‘ step whose content is the exact question that should be asked to the user when user-specific clarification is needed

    work page

  11. [14]

    Rely on your knowledge until a gap is identified that requires tool support

    Step-by-Step Approach: - Work through reasoning systematically, breaking down the task into manageable steps. Rely on your knowledge until a gap is identified that requires tool support. - Use only the domain-appropriate tool when needed

    work page

  12. [15]

    ### Output Guidelines - Your answer must begin with ‘### Reasoning‘

    Goal-Oriented Resolution: - Conclude your reasoning process by achieving a clear, accurate, and succinct solution based on your independent analysis and any tool findings. ### Output Guidelines - Your answer must begin with ‘### Reasoning‘. - After that, every step must begin with one of these section headers on its own line: ‘### Reasoning‘, ‘### Search‘...

    work page