PaperlayerView exact revision on arXiv

Source-linked AI summary

How Do Agentic LLMs Decide to Call Tools? A Tool-Call Vector Shaped by Suppression

Xijie Gong, Tingxu Han, Jiahao Zhang, Wei Song, Ziqi Ding, Hanqi Yan, Youcheng Sun, Lijie Hu

arXiv:2610.09624v1cs.AIcs.LG

TL;DR

Mechanistic analysis of tool calling is difficult because agentic prompts entangle scaffolds and requests. The paper uses single-verb contrasts to identify a causally controlling vector that transfers to native trajectories and verb-free requests.

  • Problem

    The mechanism behind tool-call versus direct-response decisions is poorly understood, while long agentic prompts make it hard to isolate a single controllable variable for mechanistic analysis.

    Source: §1 · pp. 1–2

  • Method

    The authors create matched prompts that differ only in request verb, then use mechanistic interventions to identify and test the tool-call vector µ∆.

    Source: §1 · p. 2

  • Results

    Layer-24 activation patching recovers <tool_call> on 100% of corrupt prompts, and µ∆ transfers to native multi-turn trajectories and verb-free requests.

    Source: §1 · p. 2

  • Takeaways & Limitations

    A compact, reusable representation links request semantics to the model's regulation of tool calling.

    Source: §9 · p. 10

    The primary formation analysis uses behaviorally screened coding pairs under a fixed scaffold, and broader tool-use correctness is not evaluated.

    Source: §10 · p. 10

Abstract

from arXiv · show

Tool calling, invoking external tools on demand, is central to agentic LLMs, yet the mechanism that decides whether a model calls a tool or responds directly remains poorly understood. Agentic prompts are long and heavily scaffolded, combining role instructions, tool schemas, format templates, and the user's request across hundreds of tokens, creating a noisy, highly entangled context in which no single controllable variable for mechanistic analysis is obvious. To obtain such a variable, we propose a method that converts complex agentic prompts into minimal contrastive pairs in which a single request verb determines the tool-call decision: replacing an execution-verb (e.g., \textit{write}) with an analysis-verb (e.g., \textit{discuss}) reliably flips the decision, suggesting it is mediated by a compact internal state. We construct 500 such paired prompts across Python, Java, and C++ (300 for mechanistic analysis, 200 held out for evaluation). We trace the decision to a vector, $\mu_\Delta$, that is both causally necessary and sufficient and generalizes beyond the discovery prompts to native multi-turn $\tau^2$-Bench trajectories and verb-free requests. Behavioral ablations show that the scaffold establishes a tool-call prior; Transcoder decomposition then reveals that analysis verbs suppress this prior through features signaling that tool use is unnecessary, whereas execution verbs largely leave it intact. Downstream scaffold-reading attention heads and MLP features read out the resulting state, and the same mechanism recurs across seven models from the Qwen, Mistral, and Granite families. Our code is available at https://github.com/XijieGo/MI4ToolCalling.

1 Introduction

The paper seeks a controllable way to study tool-call decisions in long agentic prompts, then uses verb contrasts to trace the decision to a compact internal state.

1 Introduction

  • In Qwen3, the decision can be analyzed as whether <tool_call> is the top-1 prediction for the first generated token.This single-token target supports activation patching and direct logit attribution.Source: §1 · p. 2
  • Mechanistic analysis is difficult because agentic prompts combine role instructions, tool schemas, format templates, and requests across hundreds of tokens.Prior tool-calling work focused on tool selection and accuracy, while prior mechanistic studies typically used short prompts with a single variable to vary.Source: §1 · p. 2
  • Replacing an execution verb with an analysis verb switches the first generated token in matched code-completion prompts that otherwise remain fixed.The study constructs 500 pairs across Python, Java, and C++, with 300 for mechanistic analysis and 200 held out for evaluation.Source: §1 · p. 2
  • The layer-24 tool-call vector µ∆ is causally necessary and sufficient, and transfers to native multi-turn trajectories and verb-free requests.Adding or subtracting µ∆ induces or suppresses tool calls, including in web retrieval, SQL execution, and email dispatch.Source: §1 · p. 2
  • Analysis verbs activate features that suppress the scaffold-induced call prior, while execution verbs largely leave it intact.Downstream scaffold-reading attention heads and MLP blocks support opening a tool-call sequence; both effects weaken when suppression features are active.Source: §1 · p. 3

2 One Verb Flips the Call-or-No-Call Decision

A single request verb can flip whether a code-completion prompt elicits a tool call or direct text. Matched prompts preserve the scaffold and task while changing only that verb.

2 One Verb Flips the Call-or-No-Call Decision

  • Replacing an execution verb with an analysis verb can switch the first output from <tool_call> to direct text.For example, asking the model to write a function often triggers a tool call, while replacing write with discuss can produce a text response.Source: §2 · p. 3
  • Matched code-completion prompts vary only the request verb, keeping task content, available tool, and scaffold fixed.Execution verbs request code action; analysis verbs request a textual response.Source: §2 · p. 3
  • Figure 2 plots first-token call rates by request type across models on HumanEval, before behavioral filtering.The caption identifies the comparison but does not report the bar values.Source: Figure 2 · p. 3
  • The Qwen3-8B dataset contains 500 prompt pairs, with 300 for mechanistic analysis and 200 held out for evaluation.Tasks span Python, Java, and C++ and draw from MBPP, APPS, HumanEval, and CodeContests.Source: §2 · p. 3
  • The study targets whether the first generated token is <tool_call> on execution prompts and not <tool_call> on their analysis-verb counterparts.The paired contrast changes only the request verb.Source: §2 · p. 4

3 The Tool-Call Vector

The tool-call decision is localized to a layer-24 residual state at the prediction position, where a vector estimated from execution–analysis differences supports call induction and suppression. This section identifies the state by patching and tests the vector’s effects on held-out prompts.

3 The Tool-Call Vector

  • The section identifies a shared residual vector by first locating the tool-call state at the final prompt position through activation patching.Other positions at the patched layer remain fixed while computation resumes from the patched state.Source: §3 · p. 4, §3.1 · p. 4

3.1 Localizing the Tool-Call State

  • At L24, patching the prediction-position state recovers calls on 100% of held-out analysis prompts.Across layers, the patching effect shifts from the verb position to the prediction position.Source: §3.1 · p. 4, Figure 3 · p. 4

3.2 Estimating the Tool-Call Vector

  • The tool-call vector is the mean execution–analysis activation difference at the L24 prediction position, estimated over 300 training pairs.The vector provides a compact causal target for tracing how components form and read out the call decision.Source: §3.2 · p. 4
  • Adding µ∆ to analysis prompts induces tool calls, while removing it from execution prompts suppresses them, showing bidirectional control at the prediction position.Both interventions recover most of the logit gap in their respective directions.Source: Table 1 · p. 4, §3.2 · p. 4

4 Generalization Beyond Verb-Cued Coding

The coding-derived direction is kept fixed at the L24 prediction position to test transfer beyond discovery prompts. The tests cover new tools and domains, native multi-turn trajectories, and requests without explicit action verbs.

4 Generalization Beyond Verb-Cued Coding

  • The tests reuse the coding-derived direction at the L24 prediction position, calibrating its norm for cross-domain experiments and using fixed gains for multi-turn and verb-free tests.Verb-free requests contain no explicit action verb specifying what the model should do.Source: §4 · p. 5, §4.3 · p. 5

4.1 Transfer Across Tools, Domains, and Action Wording

  • The coding-derived vector supports bidirectional first-token call control across web retrieval, read-only SQL execution, and email dispatch.The direction remains fixed at L24, with its norm matched to each target domain; mean normalized logit shifts were 0.88 for addition and 0.94 for removal.Source: §4.1 · p. 5, Table 2 · p. 5

4.2 Native Multi-Turn, Multi-Tool Trajectories

  • 36.7% of baseline tool calls were suppressed and 70.0% of baseline text-response turns were induced as calls in native multi-turn trajectories at 1× gain.The evaluated Telecom trajectories had long contexts, multiple available tools, and prior interaction and call history.Source: §4.2 · p. 5, Table 3 · p. 6

4.3 Verb-Free Requests

  • At 1× gain, vector removal suppressed 93.1% of baseline Qwen3-8B calls on verb-free requests.At 1.5× gain, suppression reached 100.0%, versus 16.2% for a norm-matched random vector.Source: §4.3 · pp. 5–6, Table 3 · p. 6

5 How Does the Tool-Call Vector Form?

The scaffold creates a task-conditioned tool-call prior, while analysis-active features suppress it in the formation window. Transcoder analysis and interventions trace how that suppression shapes the layer-24 tool-call vector.

5.1 The Scaffold Establishes a Tool-Call Prior

  • On held-out tasks, neutral requests average 0.8486 call probability versus 0.0038 for analysis requests; removing the format template nearly eliminates calls.With the format template alone, call probabilities approach one for both request types.Source: §5.1 · p. 6, Table 4 · p. 6
  • The prior is task-conditioned: execution requests reach 100% calls, analysis requests 0%, and empty or unrelated turns elicit none.The task body alone yields a 32.3% call rate.Source: §5.1 · p. 6

5.2 MLPs Dominate Tool-Call Vector Formation

  • In L20–L23, MLPs supply 77.7% of projected formation-window writes versus 22.3% from attention heads; MLP23 contributes the largest single write, 22.5 on the g_l scale.The execution–analysis gap rises sharply in this window after remaining near zero through the first fifteen layers.Source: §5.2 · p. 7, Figure 4 · p. 7

5.3 Suppression Shapes the Vector in the Formation Window

  • Transcoders approximate MLP outputs with feature-weighted decoder directions; feature selection uses 300 training pairs, with scoring and interventions on 200 held-out pairs.Features are scored by activation contrast and projection onto ˆµ∆.Source: §5.3 · p. 7
  • Analysis-active features dominate L21–L23: the 115-feature analysis-non-execution family has mean held-out |κ| 0.0633 versus 0.0156 for the 292-feature execution-request family.The analysis-non-execution family totals 7.28 of the full-feature total 14.34; the window-level K_corrupt/K_clean ratio is 1.99.Source: §5.3 · p. 8, Table 5 · p. 8
  • Zeroing five selected suppressors raises held-out call rate from 3.0% to 25.5%, with 22.5% strict recovery (45/200), versus 0.5% recovery for the random control.The intervention shifts the state +5.30 along ˆµ∆, closes 8.82% of the gap, and raises mean call margin by +1.38.Source: §5.3 · p. 8
  • Projecting onto ˆµ∆ at L24 recovers the baseline gap and all 200 clean calls, supporting the vector’s use in the feature interventions.The analysis-side suppressors mostly activate on analysis prompts and write against ˆµ∆.Source: §5.3 · p. 8

6 How Does the Tool-Call Vector Control the Call Decision?

Downstream components read the L24 tool-call state into the first generated token. Scaffold-reading attention heads and a late structural MLP feature both increase support for the tool-call opening.

6 How Does the Tool-Call Vector Control the Call Decision?

  • Direct logit attribution and single-component patching over L25–L35 trace how downstream attention heads and MLP features convey the L24 state to the first token.Direct logit attribution measures each component’s direct contribution to the tool-call logit.Source: §6 · p. 8, §6.2 · p. 9

6.1 Scaffold-Reading Heads Support Tool Calling

  • Adding µ∆ to analysis prompts raises scaffold-reading heads’ direct support for <tool_call>; L33H29’s DLA rises from 11.13 to 24.43, near its execution baseline.L29H9, L33H11, and L33H29 shift attention toward the tool-format template and away from role instructions on execution prompts.Source: §6.1 · p. 8, Figure 5 · p. 9

6.2 Committing to the Tool-Call Opening

  • Patching MLP34 yields a 41.0% call rate; replacing F109925’s analysis-side activation yields 37.0%, versus 6.5% for the non-structural control F91365.F109925 responds to tool-schema boundaries and writes toward the tool-call opening; its effect is smaller than the full MLP34 patch.Source: §6.2 · p. 8, Figure 5 · p. 9

7 The Tool-Call Vector Across Scales and Model Families

Across seven Qwen, Mistral, and Granite models, the tool-call mechanism retains strong localization and causal intervention, while generalization and downstream readout vary by architecture. The analysis-side suppression asymmetry persists across the tested models.

7 The Tool-Call Vector Across Scales and Model Families

  • The cross-model tests include Qwen3-4B and 14B, Qwen3.5-4B and 9B, Mistral-Small-3.2-24B, and Granite-3.3-8B.The study assesses localization, causal interventions, generalization, formation-window writes, and scaffold attention.Source: §7 · p. 9, Table 6 · p. 9
  • Across all seven models, full-state patching recovers calls on 93.5–100.0% of analysis prompts; vector sufficiency scores are 0.81–1.03 and necessity scores 0.61–1.04.The table summarizes recurrence and generalization beyond discovery prompts across models and intervention axes.Source: §7 · pp. 9–10, Table 6 · p. 9
  • K_corrupt exceeds K_clean in every model, preserving the finding that the vector is shaped primarily by analysis-side suppression rather than an execution-specific signal.The specific layers and components vary across models, while the functional organization remains stable across those tested.Source: §7 · p. 10

8 Related Work

Prior work evaluates tool-use decisions and studies mechanistic representations, but the causal internal representation of tool initiation remains largely uncharacterized. This paper connects those lines through controlled contrasts and a localized tool-call vector.

8 Related Work

  • Tool-use research evaluates whether and how models invoke tools, while the causal internal representation of that decision remains largely uncharacterized.Mechanistic interpretability has localized causal states and resolved component computations; representation engineering has studied compact residual directions mediating behavior.Source: §8 · p. 10
  • The paper extends causal mechanistic analysis to tool initiation by using controlled request contrasts in long, scaffolded prompts.This connects mechanistic interpretability with tool-use evaluation and work on compact residual directions.Source: §8 · p. 10
  • The analysis traces how request semantics suppress a scaffold-induced call prior, linking tool-use behavior to an internal causal account of action initiation.The reported vector transfers across tools, domains, and request forms.Source: §8 · p. 10

9 Conclusions

A compact tool-call vector controls the first-token call decision in Qwen3-8B and transfers to native multi-turn and verb-free requests. The paper describes a scaffold-induced call prior suppressed by analysis requests and read out by downstream components.

9 Conclusions

  • A single tool-call vector gives causal control over Qwen3-8B’s call-or-no-call decision at the first output token.The vector transfers without re-estimation to native multi-turn trajectories and verb-free requests.Source: §9 · p. 10
  • The scaffold establishes a call prior on task-bearing requests, while analysis requests activate suppressor features and execution requests largely preserve that prior.Downstream attention and MLP features read out the resulting state into the tool-call opening.Source: §9 · p. 10
  • The tool-call vector connects these computations through a compact, reusable representation of how request semantics regulate tool calling.Source: §9 · p. 10

10 Limitations

The paper limits its claims to the binary call-or-no-call mechanism, studied primarily with behaviorally screened coding pairs under fixed scaffolds. Its controlled pairs vary request verbs while preserving task context and tool interface, and the layer-24 state supports bidirectional control.

10 Limitations

  • The study addresses the binary call-or-no-call decision, not broader tool-use correctness, and its fixed-scaffold coding-pair findings may not extend to other prompt constructions.The formation analysis does not resolve the fine-grained computation mapping scaffold information and request semantics to the vector.Source: §10 · p. 10

A Experimental Setup and Dataset Construction

A.1 Source Tasks

  • The model-specific corpora draw on programming tasks or τ^2-Bench Telecom conversations, with source families varying by model.Coding-task normalization preserves function behavior, programming language, and available examples or tests.Source: §A.1 · p. 16, Table 7 · p. 16

A.2 Agentic Prompt Template

  • Qwen3-8B prompts use a native chat template with a write_file tool and specified <tool_call> serialization, while cross-model corpora use each model’s native template and call-opening token.Within a pair, the tool schema, output format, task body, and assistant prefix remain identical.Source: §A.2 · pp. 16–17

A.3 Pair Construction Pipeline

  • Each model has 500 retained pairs, split into 300 training and 200 held-out pairs, with vector fitting on training data and interventions evaluated on held-out pairs.Source: §A.3 · p. 17, Table 8 · p. 17, §A.4 · p. 17

B.1 Layer and Position Controls

  • Addition induces calling on all evaluated analysis prompts from L23 onward; removal suppresses every evaluated execution-side call at L24 and L25, where L24 combines high full-state recovery in the localization sweep with complete bidirectional control.The neighboring-layer check fits mean-difference directions at L22–L25 using 200 training pairs and evaluates them on 300 disjoint held-out pairs from a separate control corpus, comparing each against an equal-norm random direction.Source: §B.1 · pp. 18–19

C Generalization and Transfer

  • Generalization tests keep the fitted direction fixed while changing task domain, tool schema, context, or wording; target-domain norm matching changes scale, not direction.Source: C Generalization and Transfer · p. 19

C.1 τ 2-Bench: Full Cross-Model Results

Across seven models, the fitted direction outperforms norm-matched random controls on both suppression and induction in native τ^2-Bench decisions, and also suppresses calls on verb-free requests. The formation and readout analyses support a context-dependent suppression account in which request semantics and tool affordances shape a call state that downstream components express.

C.1 τ 2-Bench: Full Cross-Model Results

  • The fitted direction exceeds norm-matched random controls for both removal and induction in every model on native τ^2-Bench decisions.The removal and induction flip rates are measured among baseline tool-call and baseline text decisions, respectively.Source: §C.1 · pp. 19–20, Table 12 · p. 19, Table 13 · p. 19

C.2 Implicit, Verb-Free Requests: Full Cross-Model Results

  • On verb-free requests, the fitted direction suppresses more baseline calls than the random control in all seven models.Source: §C.2 · p. 20, Table 14 · p. 20

D.1 Scaffold and Request Baselines

  • Scaffold components and request wording establish a conditional call prior: neutral and execution requests preserve high call probability, while analysis requests strongly reduce it.A length-matched schema control places the neutral–analysis gap between the full-scaffold and no-schema conditions, implicating both schema semantics and context length.Source: §D.1 · pp. 20–21, Table 15 · p. 20

D.2 Request–Tool Affordance Alignment

  • Matched request–tool pairs both reach 100% calling, including Review paired with a review-submission tool, supporting a decision that combines request semantics with tool affordances.Source: §D.2 · p. 21, Table 17 · p. 21

D.3 Early Separability vs. Late Causal Commitment

  • The clean/corrupt distinction is linearly recoverable at L1, but the logit-lens gap first exceeds 1.0 at L27, separating early signal from later causal commitment.The largest positive increments accumulate late in the stack.Source: §D.3 · p. 22, Table 18 · p. 22

D.4 Feature-Level Formation Summary

  • The analysis-non-execution feature family has about four times the mean held-out |κ| of the execution-request family, and 102 of its 115 training-labelled features write against the tool-call direction.Source: §D.4 · p. 22, Table 19 · p. 22

D.5 Suppression Accumulates Across the Formation Window

  • Suppressor-feature writes shift the L24 input state by +5.30 along the inverse tool-call direction and raise the held-out call rate from 0.0% to 25.5%, versus 0.5% strict recovery for the random control.The formation window includes substantial suppression contributions before L24, with L23 contributing 55.9% of net feature write.Source: §D.5 · pp. 22–23, Table 20 · p. 23, Table 21 · p. 23

E Reading Out the Tool-Call Vector

  • The evidence favors context-dependent suppression and distributed downstream readout: analysis-side features suppress the prior, while scaffold-reading heads and MLPs contribute to expressing the call state.L25–L29 heads often show high single-head recovery, whereas L33 heads show larger DLA; MLP34 has the highest single-block recovery in the component sweep.Source: §D.6 · p. 24, §E.1 · pp. 24–25, E Reading Out the Tool-Call Vector · p. 24

E.3 MLP34 Structural Readout

A late MLP feature links the vector-controlled decision state to the tool-call serialization format, while downstream readout remains distributed rather than a strict serial path. The cross-model section tests localization and shared-direction effects across architectures.

E.3 MLP34 Structural Readout

  • L34 feature F109925 responds to tool-schema boundary contexts, projects positively onto the <tool_call> token, and approaches its clean activation and write after vector addition.Replacing its analysis-side activation increases the mean call-token logit by 0.56 and margin by 0.92.Source: §E.3 · pp. 25–26, Table 25 · p. 25, Table 26 · p. 26

E.4 Alternative Downstream Path Controls

  • Removing selected early scaffold-reading heads leaves late DLA largely intact, while removing the L24 vector reduces late-layer DLA and group ablations support a distributed set of downstream writers.The combined group intervention produces a 5.02 logit drop, 2.16 times the best single-module result, while flipping 5/500 clean prompts.Source: §E.4 · p. 26

F Cross-Model Validation

F.2 Cross-Model Early Separability vs. Late Commitment

This section examines cross-model timing and readout of tool-call information, with measurements across seven models supporting early linear separability and later output-space expression and direct layer writes. It also reports the experimental scope and compute setup, including a sparse forward-core replay diagnostic with Qwen3-1.7B.

F.2 Cross-Model Early Separability vs. Late Commitment

  • Probe AUC exceeds 0.9 at L0–L1 in every model, while the logit-lens gap exceeds 1.0 at L14–L27 and the largest positive DLA writes occur in upper layers.The DLA layers are L34–L39 in 40-layer models, L29–L34 in 36-layer models, and L23–L31 in 32-layer models; causal commitment at L∗ is a separate measurement. Some models show large negative final-layer writes, which are excluded from positive-writer ranking.Source: §F.2 · pp. 27, 29, Figure 6 · p. 28

F.3 Cross-Model Formation and Readout Summary

  • Every model has greater corrupt-higher than clean-higher feature mass, while scaffold readers shift attention toward format regions in Qwen models and tool schemas in selected Mistral and Granite heads.Reader identities and scaffold regions vary by architecture, but attention changes with call state and selects scaffold content relevant to output format.Source: §F.3 · p. 29

G.1 ACDC

  • At 8B and 14B, ACDC cores have near-unit sufficiency, but size-matched random cores also replay strongly, with gaps of 0.0232 and 0.0446.High sufficiency therefore offers limited evidence for a uniquely selected route; the vector intervention instead tests the call state at a localized residual coordinate.Source: §G.1 · p. 30, Table 31 · p. 30

G.2 EAP-IG

  • At 10,000, 50,000, and 100,000 retained edges, EAP-IG mean sufficiency scores are −5.19, 3.77, and 9.25, respectively.The clean baseline is 14.50; 50,000 edges induce 937 nodes and 100,000 induce 1,075, while clean-side removal scores are −13.06, −13.00, and −13.38 at those edge budgets. Retained nodes cover the 18-node forward core from 2,000 edges onward, while faithfulness increases with edge budget.Source: §G.2 · p. 30, Table 32 · p. 31

G.4 Analysis Granularity

  • Circuit and feature diagnostics describe distributed computation, while the L24 intervention summarizes the call decision with a single residual direction and directly tests its causal effect.Activation patching localizes the prediction-position state, direction interventions test necessity and sufficiency, and feature and attention measurements explain its formation and downstream readout.Source: §G.3 · p. 31, §G.4 · p. 31, Figure 8 · p. 32

H Related Work

  • Prior tool-use work largely treats invocation as an external behavioral outcome; this paper presents a causal account of action initiation, combining representation-, circuit-, and feature-level analysis.The account uses representation-level localization, circuit diagnostics, and Transcoder decomposition to analyze the tool-call state and its contributors.Source: §H.1 · p. 33, §H.2 · pp. 33–34

H.3 Behavior-Mediating Directions and Steering Vectors

  • The scaffold establishes an elevated call prior, while analysis requests activate suppressor features; local removal of µ∆ at layer 24 reduces clean top-1 call rate from 100.00% to 0.00%.Global direction ablation leaves a 40.04% clean call rate; the fixed coding-derived direction also transfers to τ²-Bench and verb-free tests without re-estimation.Source: §H.3 · p. 34, §H.4 · p. 35

I Implementation and Reproducibility

  • The evaluation uses held-out inputs and native call-opening tokens for cross-model attention measurements, and the release includes experiment code, paired datasets, and Transcoder checkpoints.The code and experiment instructions are available at https://github.com/XijieGo/MI4ToolCalling.Source: I Implementation and Reproducibility · p. 35, §I.1 · p. 35, §I.2 · p. 35

I.3 Compute Resources

  • Training the Transcoders required approximately 800 NVIDIA B200 GPU-hours; other reported experiments can be reproduced on one RTX PRO 6000 Blackwell GPU with 96 GB memory.The total includes debugging and production training; reproduction assumes the provided checkpoints.Source: §I.3 · p. 36

I.4 LLM Usage

  • GPT-5 generated debugging examples; primary mechanistic experiments evaluated seven listed Qwen, Mistral, and Granite models, with Qwen3-1.7B included in an additional sparse forward-core replay diagnostic.The primary models were Qwen3-4B, Qwen3-8B, Qwen3-14B, Qwen3.5-4B, Qwen3.5-9B, Mistral-Small-3.2-24B, and Granite-3.3-8B.Source: §I.4 · p. 36

Related papers

Loading 2610.09624v1…