Cognitive Diversity in LLM Tool-Use: Behavioural Fingerprints, Convention Adherence, and the Case for Substrate Mixing

1. Introduction

The selection of a language model for an agentic system is typically driven by benchmark performance: pass rates on coding tasks, accuracy on question-answering, throughput per dollar. These metrics answer the question how well does this model perform? but leave a more consequential question unanswered: how does this model think?

Two agents can achieve the same pass rate while exhibiting radically different cognitive strategies. One discovers relevant files through pattern search; another navigates by memorised paths. One refuses to produce a specification when an adequate one already exists; another writes a fresh one regardless. One reflects on its own blind spots with calibrated confidence; another treats its output as self-evidently correct. These differences are invisible to accuracy-based evaluation, yet they determine what an agent sees, what it misses, and whether its output is safe to act on without human review.

We call these stable, lineage-specific patterns cognitive fingerprints. They are not bugs or failures — they are the natural consequence of different training data, architectures, and alignment procedures producing different cognitive phenotypes. The practical question is not whether fingerprints exist (they manifestly do) but whether they matter for system design, and if so, how to measure and exploit them.

The monoculture problem

Most agentic systems deploy a single substrate. This creates a monoculture: a systematic blind spot that no amount of prompt engineering can eliminate, because the blindness is architectural rather than instructional. If your substrate discovers files through Grep but never uses Glob, it will find cross-file references but miss structural patterns visible only through directory traversal. If it resolves ambiguity by deferring to existing specifications rather than proposing alternatives, it will never surface the creative solutions that come from treating each task as fresh.

The multi-agent literature has begun to address this. The X-MAS framework (Zhu et al., 2025) demonstrated a 47% improvement on mathematical reasoning by mixing chatbot and reasoner architectures. But mixing for cognitive diversity — deliberately selecting substrates with complementary fingerprints — remains unexplored.

Convention adherence: the missing axis

Current evaluation frameworks measure whether an agent follows its instructions (IFEval, AGENTIF, FireBench). But instructions and conventions are not the same thing. An instruction is given to the agent explicitly: “write a work package.” A convention is a norm the agent must discover and internalise: “work package numbers must not collide with existing ones.” The distinction matters because instruction following is a compliance test, while convention adherence is a cultural competence test.

In our experiment, every substrate that attempted to assign a work package number chose the same wrong number — one already claimed by an existing work package. The convention for checking available numbers was documented in the project, and every substrate had the tools to verify it. None did. This universal failure is invisible to instruction-following benchmarks because the instruction was followed correctly (a number was assigned); the convention was violated (it was the wrong number).

What this survey contributes

We present a five-dimensional evaluation framework and apply it to thirteen substrates performing the same specification-authoring task:

1. Behavioural fingerprinting — tool-use distributions as cognitive signatures
2. Tool-use strategy — how foraging patterns determine what gets discovered
3. Substrate diversity — complementary coverage through lineage mixing
4. Beyond-accuracy evaluation — six qualitative axes replacing pass/fail
5. Convention adherence — whether agents absorb organisational norms they discover, not just instructions they receive

No existing study combines these five dimensions in a single experimental frame. We ground each dimension in its literature, present our case study, and identify the gap this work fills.

2. Related Work

2.1 Behavioural Fingerprinting

The idea that language models have stable behavioural profiles has been explored through two lenses: personality psychometrics and provenance fingerprinting.

Pei et al. (2025) introduced a Behavioral Fingerprinting framework using a Diagnostic Prompt Suite to profile eighteen models across capability tiers. Their finding that “core capabilities like abstract and causal reasoning are converging among top models, [while] alignment-related behaviors such as sycophancy and semantic robustness vary dramatically” supports our observation that fingerprints are most visible in how models approach tasks, not whether they solve them. They also documented a cross-model default persona clustering (ISTJ/ESTJ) that likely reflects common alignment incentives.

The Nature Machine Intelligence framework (2025) applied psychometric validation to eighteen LLMs, finding that personality measurements in instruction-tuned models are reliable and valid under specific prompting configurations. This confirms that cognitive style is not noise — it is a measurable property of the substrate.

A separate line of work uses behavioural patterns for provenance tracking rather than cognitive profiling. The Refusal Vectors approach (arXiv 2602.09434) leverages directional patterns in internal representations when processing harmful versus harmless prompts. AgentPrint achieves F1=0.866 in agent identification through traffic fingerprinting of tool-use patterns. These provenance methods confirm that tool-use behaviour is distinctive enough to serve as an identifier — precisely the property we exploit for cognitive profiling.

What the fingerprinting literature lacks is a controlled comparison of fingerprints on the same task. Studies profile models individually or compare them on distinct benchmarks. Our experiment profiles thirteen models on a single task with identical tool access, making the fingerprints directly comparable.

2.2 Tool-Use Benchmarking

The tool-use evaluation landscape has matured rapidly since ToolBench (Qin et al., 2024; ICLR 2024), which established the methodology for testing LLM agents with real API calls. Three recent benchmarks extend this work:

MCP-Bench (Ding et al., 2025; arXiv 2508.20453) connects LLMs to 28 live MCP servers spanning 250 tools. Unlike prior API-based benchmarks, each server provides complementary tools designed for multi-step coordination. MCPAgentBench (arXiv 2512.24565) extends this to real-world tasks.

BFCL V4 (Berkeley Function Calling Leaderboard) evaluates serial and parallel function calls across programming languages using AST-based evaluation, scaling to thousands of functions.

The Springer survey (Xu et al., 2025) provides a systematic review of tool-learning agents, covering retrieval, planning, and emerging frontiers including multimodal tools.

These benchmarks measure tool-use competence — can the model call the right tool with the right parameters? Our experiment measures tool-use strategy — which tools does the model choose when multiple are available and equally valid? This is the difference between whether a craftsperson can use a chisel and which tools they reach for by habit.

2.3 Multi-Agent Diversity

X-MAS (Zhu et al., 2025; arXiv 2505.16997) is the closest work to our substrate-mixing argument. Their X-MAS-Bench evaluates 27 LLMs across 5 domains and 5 functions (1.7 million evaluations), and X-MAS-Design demonstrates that heterogeneous agent combinations (chatbot + reasoner) consistently outperform homogeneous systems. The 47% improvement on AIME-2024 is striking evidence for complementary substrate selection.

Intrinsic Memory Agents (Yang et al., 2025; arXiv 2508.08997) address how agent-specific memories evolve intrinsically with agent outputs rather than through external summarisation. The framework maintains role-aligned memory that preserves specialised perspectives — a mechanism for retaining cognitive diversity within a multi-agent system rather than homogenising it.

MultiAgentBench (Zhu et al., 2025; ACL 2025; arXiv 2503.01935) evaluates collaboration and competition across coordination protocols (star, chain, tree, graph topologies). Their finding that cognitive planning improves milestone achievement by 3% and that graph structure performs best in research scenarios suggests that diversity benefits depend on communication topology.

Talebirad & Nadiri (2023) proposed an early framework for multi-agent collaboration with LLMs, demonstrating case studies in Auto-GPT and BabyAGI. While foundational, the work predates the current generation of tool-using agents and does not address substrate diversity.

What the diversity literature lacks is a principled method for selecting which substrates to mix. X-MAS shows that mixing helps; our work shows which cognitive fingerprints complement each other — Grep-heavy substrates find cross-file references that Read-only substrates miss, and vice versa.

2.4 Beyond-Accuracy Evaluation

SWE-Bench Pro (Deng et al., 2025; arXiv 2509.16941) extends SWE-bench to long-horizon enterprise tasks across 41 repositories, evaluating maintainability, readability, and security alongside pass/fail. The best models reach only 35.3% success on complex tasks, confirming that accuracy alone is an insufficient measure.

GAIA evaluates agents on real-world tasks requiring tool use, multi-step reasoning, and information retrieval. The highest score at end of 2025 was 90%, suggesting that for well-defined tasks, top models approach ceiling. The interesting variation is how they reach that ceiling.

AgentArch (Bogavelli et al., 2025; arXiv 2509.10769) is the first benchmark systematically evaluating 18 agentic architectures across 6 LLMs on enterprise workflows. Their key finding — that “optimal configurations vary by model and task complexity, rather than following a single best-performing design” — directly supports our substrate fingerprinting thesis.

The CLASSic framework (Aisera, 2025) proposes five evaluation dimensions (Cost, Latency, Accuracy, Stability, Security) with empirical evidence that domain-specific agents achieve 82.7% accuracy versus 59–63% for general LLMs at 4.4–10.8× lower cost. Beyond-accuracy evaluation is becoming a practical requirement, not an academic exercise.

The agent evaluation survey (arXiv 2507.21504) identified 120 evaluation frameworks and flagged missing enterprise requirements: multistep granular evaluation, cost-efficiency measurement, safety and compliance focus, and live adaptive benchmarks. Our six-dimensional framework addresses several of these gaps.

2.5 Convention Adherence

The Instruction Gap (Tripathi et al., 2025; arXiv 2601.03269) tested 13 LLMs on instruction compliance in RAG scenarios, finding that models “excel at general tasks but struggle with precise instruction adherence.” Claude Sonnet 4 and GPT-5 achieved the highest results. This aligns with our finding that Anthropic substrates excel at convention adherence, but extends it: our experiment tests discovered conventions, not given instructions.

AGENTIF (2025; arXiv 2505.16944; NeurIPS 2025) is the first benchmark for agentic instruction following, featuring 50 real-world applications with instructions averaging 1,723 words and 11.9 constraints each. The best model perfectly follows fewer than 30% of instructions — a sobering baseline for convention adherence.

FireBench (arXiv 2603.04857; March 2026) evaluates six capability dimensions across enterprise applications including format compliance, ranked responses, and mandatory inclusions/exclusions. This is the closest benchmark to convention adherence, but still tests explicit instructions rather than discovered norms.

IFEval remains the most widely used instruction-following benchmark, with its strength in formalising multi-constraint compliance. However, its synthetically constructed instructions (average 45 words) are far simpler than the conventions agents encounter in real projects.

2.6 The Gap

No existing study:

• Tests 10+ substrates on the same task with same tools
• Treats tool-call patterns as cognitive fingerprints (not just success rates)
• Measures convention adherence (not instruction compliance)
• Maps convergence/divergence boundaries across substrates
• Combines all five dimensions in a single experimental frame

This survey fills that gap.

3. Case Study: MāyāLucIA

MāyāLucIA is a human-machine collaborative intelligence project organised around spirits (named agents with persistent identity), guilds (domain collectives), and a relay (append-only broadcast for coordination). The project uses work packages (WPs) as its unit of specification — each WP is a self-contained briefing for agent execution with standardised sections: genesis, context, inventory, specification, execution order, acceptance criteria.

The spirit registry (“aburaya”) maintains identity files for each agent, cross-referenced with guild membership, exported powers (LLM-intelligible procedures), and project assignments. This registry is the subject of the audit task used in our experiment.

Why this project?

MāyāLucIA provides a controlled experimental setting because:

1. Rich cross-reference structure — the registry contains deliberate gaps (spirits without guilds, phantom references, stale documentation) alongside working components. An audit task has ground truth.

2. Established conventions — work package authoring follows a documented convention with specific structural requirements. Convention adherence can be measured against a known standard.

3. Multi-substrate history — the project has been developed across multiple substrates, providing baseline expectations for how different lineages interact with the same codebase.

4. Methods

4.1 Task Design

The experiment uses the author-wp power — an LLM-intelligible procedure that instructs an agent to:

1. Survey the spirit registry for cross-reference gaps (Phase 1)
2. Author a work package specifying repairs (Phase 2)
3. Reflect on the reproducibility and confidence of its own output

The task combines exploration (discovering gaps through file reading and search), specification (translating discoveries into an actionable WP), and metacognition (assessing the quality and reproducibility of its own work). This three-phase structure separates survey competence from specification competence from self-awareness.

4.2 Substrate Selection

Thirteen substrates from nine lineages:

Three substrates were tested in Phase 1 (Kimi K2.5, Gemini, Qwen) with manual orchestration. Ten additional substrates were tested in Phase 2 using the gaddi orchestrator — a hook-based automation system that fires prompts after each response completes, with zero confirmation gates.

4.3 Tool Configuration

All substrates received identical read-only tools:

The read-only constraint ensures substrates cannot modify the registry during audit. Tool configuration is a control-plane variable: the same substrate produces different autonomy outcomes depending on which tool registry the session uses.

4.4 Evaluation Dimensions

We evaluate six dimensions, none of which are captured by standard pass/fail metrics:

1. Survey completeness — how many of the known gaps were discovered? Ground truth: 10 distinct finding classes established by union of all substrate outputs.

2. Specification quality — is the WP actionable? Measured by: exact file paths, before/after diffs, testable acceptance criteria.

3. Convention adherence — does the WP follow the project’s established structure? Sections present, WP number validity, executor choice, relay announcement.

4. Interpretive divergence — where substrates disagree, what clustering patterns emerge? This maps the boundary between automatable and judgment-requiring decisions.

5. Reflection quality — four tiers from absent to meta-theoretical. Does the substrate assess its own confidence, identify its own blind spots, predict how other substrates would differ?

6. Cost efficiency — dollars per unique finding class. Not total cost, but discovery cost per gap.

4.5 Orchestration

Phase 1 (3 substrates) used manual orchestration via emacsclient injection into gptel (an Emacs-based LLM client). This approach suffered from streaming corruption when background Emacs processes injected messages into the API response stream.

Phase 2 (10 substrates) used the gaddi orchestrator — a buffer-local prompt queue hooked into gptel’s response completion system. The gaddi waits for the terminal FSM state (DONE/ERRS/ABRT) before injecting the next prompt, eliminating the streaming corruption that plagued Phase 1. The gaddi ran ten substrates with zero manual intervention.

5. Results

5.1 Tool-Foraging Fingerprints

Tool-use distributions reveal four distinct foraging strategies:

Four strategy clusters emerge:

1. Read-only (Grok): navigates entirely by file paths, no discovery phase. 23 Read calls, zero search.
2. Read-dominant (GPT-5.2, Haiku, DeepSeek, GLM-5, MiniMax, Kimi K2T, Sonnet): the majority strategy. Read 56–77% of tool calls.
3. Grep-heavy (Opus): discovers through pattern search — 35% of tool calls are Grep. The cross-reference hunter.
4. Glob-heavy (Step): discovers through directory listing — 44% Glob, zero Grep. The structural surveyor.

These patterns are lineage-stable. The three Anthropic models (Opus, Sonnet, Haiku) show a family gradient: Opus is Grep-heavy (35%), Sonnet is balanced (10% Grep, 20% Todo), Haiku is Read-heavy (77% Read). The two Moonshot models (K2.5 and K2 Thinking) both show Read-dominant patterns with moderate Glob.

5.2 Survey Depth vs Specification Quality

An initial hypothesis — that survey depth and specification quality are inversely correlated — was disproved by Phase 2 data. The apparent trade-off in Phase 1 (Gemini found 12 gaps but no WP; Kimi found 4 gaps with a tight WP) was an artefact of context management, not a fundamental cognitive constraint. When substrates have enough interaction turns to separate “explore” from “specify” temporally, the inverse correlation disappears. Haiku found ~8 gaps AND produced a comprehensive WP. GPT-5.2 found 5 gaps AND traced through validator source code to predict exact failure points.

5.3 Convergence and Divergence Boundaries

The convergence matrix maps which findings each substrate discovered:

The boundary: findings with >85% agreement are mechanical — any competent substrate will find them. Findings with 23–38% agreement require specific foraging strategies (Grep-heavy or deep-reading) and are systematically missed by substrates with limited search behaviour. The deep findings (identity drift, power cross-refs, unclaimed powers) were found only by the two largest-context substrates (Gemini, Opus) and the Anthropic family (Sonnet, Haiku).

This pattern provides a general method: run N substrates on the same task. Where >85% agree, automate. Where they diverge into 3+ clusters, escalate to human judgment. The boundary itself is the finding.

5.4 Convention Adherence

Two convention failures were universal or near-universal:

WP number collision: every substrate that produced a WP chose number 0042 — already assigned to an existing work package. None checked the project’s workpacks/ directory for existing numbers. The convention for number assignment was documented; the tools to verify it were available; no substrate used them. This is a convention discovery failure, not an instruction-following failure.

WP-refusal (lineage-specific): Claude Opus 4.6 and Claude Sonnet 4.5 independently declined to write a new WP after discovering that an existing WP (0041) was a superset of what they would have produced. Both produced meta-WPs instead — verification reports and tightening recommendations. No non-Anthropic substrate exhibited this behaviour. Claude Haiku 4.5 did not refuse — it produced a fresh WP positioned as consolidating and superseding prior work.

The WP-refusal pattern is governance-aware judgment: the Anthropic substrates inferred a norm (“don’t duplicate specifications”) from the project’s existing WP lifecycle and supersession mechanics. Whether this is a strength (avoiding specification sprawl) or a weakness (non-compliance with the explicit task) depends on what the experiment measures. We consider it a finding, not a failure.

5.5 Interpretive Divergence: Guild Assignment

The key substrate-dependent decision was what guild to assign to two guildless spirits:

*DS and GPT with nuance: DS deferred to open questions; GPT assigned different guilds per spirit (mayadev→mayalucia, mu2tau→apprentis).

Three clusters, each with a defensible rationale. Cluster 1 maps to the existing organisational vocabulary. Cluster 2 introduces a new concept (trans-guild) absent from the project’s glossary — creative but potentially destabilising. Cluster 3 recognises the decision as requiring human judgment and refuses to assume.

GPT-5.2 was unique in applying per-spirit semantic reasoning rather than a uniform rule — the most nuanced approach, and the only one that distinguished the two spirits’ different organisational roles.

5.6 Reflection Quality

Four tiers emerged:

Tier 1 — exceptional self-critique (Opus, Sonnet, Haiku, Qwen): Three-tier reproducibility assessment (mechanical/judgment/argued), calibrated confidence percentages, meta-observation on scope as judgment, discovery/interpretation boundary explicitly named.

Tier 2 — structured and honest (Gemini, GPT-5.2, Grok): Acknowledged own failures (Gemini: “I never actually wrote the WP”), confidence bands per criterion, Known/Inferred/Speculated taxonomy.

Tier 3 — adequate with blind spots (Kimi K2.5, GLM-5, Kimi K2T, MiniMax, DeepSeek): Structured but missed own errors (K2T didn’t catch its mu2tau miss; MiniMax didn’t catch its internal contradiction).

Tier 4 — insufficient (Step 3.5 Flash): Did not acknowledge that no specific gaps were found; treated taxonomy as equivalent to data-driven audit.

Reflection quality correlates with survey depth (more files read → more material for self-critique) but is independent of specification quality. The Anthropic lineage occupies 3 of 4 Tier 1 positions. Whether this reflects a lineage trait or a confound (the task preamble was written by an Anthropic substrate) is an open question.

5.7 Cost and Efficiency

*Step: taxonomy from inference only, zero empirical gaps.

Cost does not predict quality linearly. The cheapest substrate producing a good WP (Grok, $0.046) costs 58× less than the most expensive (Opus, $2.70), yet Opus found 2.4× more gaps. The practical question is: what gap coverage do you need? For mechanical cross-reference checks, $0.05 suffices. For deep structural audits, >$0.36 is required.

Gap discovery cost ($/gap) is a more useful metric than total cost. By this measure, GLM-5 ($0.003/gap) and Grok ($0.009/gap) are the efficiency leaders. Opus ($0.225/gap) is 75× more expensive per gap — but finds gaps the cheap substrates structurally cannot reach.

5.8 Token Metrics

6. Discussion

6.1 Monoculture Blindness as Practical Risk

Our convergence boundary analysis quantifies the cost of monoculture. A system using only Grok (the cheapest effective substrate) would find 5 of 10 finding classes — a 50% blind spot rate. A system using only Opus (the most expensive) would find 12 — but at 58× the cost. A two-substrate system (Grok + Opus) would find 13 of 13 at a combined cost of $2.75.

The practical recommendation is not “use the most expensive model” but “use complementary fingerprints.” Grok’s Read-only strategy finds path-inferable gaps. Opus’s Grep-heavy strategy finds cross-reference gaps. Neither finds what the other does. Together, they achieve complete coverage.

6.2 Tool-Use as Cognitive Phenotype

We propose treating tool-use distributions as cognitive phenotypes — stable, measurable properties of a substrate that predict its discovery capabilities. Like biological phenotypes, cognitive phenotypes:

• Are lineage-specific (the three Anthropic models show a family gradient)
• Are task-independent (the same foraging strategy appears across different tasks)
• Determine what the organism can perceive (Grep-heavy substrates see cross-file references; Read-only substrates see what their path knowledge permits)

This framing moves beyond the “which model is best?” question to “which model sees what?” — a fundamentally different evaluation paradigm.

6.3 When to Mix and When Not To

Not all tasks benefit from substrate mixing. Our data suggests:

Mix when: the task has a large search space (many files, many possible gaps), when convention adherence matters, when interpretive divergence is expected (design decisions, not mechanical repairs).

Don’t mix when: the task is well-defined with clear acceptance criteria, when cost is the binding constraint and coverage is acceptable, when speed matters more than completeness.

The convergence boundary provides a decision rule: run a small N-substrate pilot. If agreement is >85%, a single substrate suffices. If agreement is <50%, mixing is required.

6.4 Limitations

1. Single project: our findings are validated on one codebase. The MāyāLucIA registry is richly cross-referenced but may not represent all agentic system architectures.

2. Single task class: specification authoring. Different task types (debugging, refactoring, testing) may produce different fingerprint patterns.

3. Tool-use only: our experiment provides read-only tools. Substrates with write access might show different foraging strategies.

4. No causal claims: we observe correlations between foraging strategy and gap discovery. We do not prove that strategy causes different outcomes — compensatory mechanisms may exist.

5. Preamble bias: the task preamble (orient-to-mayalucia power) was authored by an Anthropic substrate. This may introduce a confound favouring Anthropic models in convention adherence and reflection quality.

7. Conclusions and Future Directions

Toward a taxonomy of substrate cognitive styles

Our four-cluster foraging typology (Read-only, Read-dominant, Grep-heavy, Glob-heavy) is a first step toward a cognitive style taxonomy. A richer taxonomy would incorporate planning behaviour (TodoWrite usage), specification structure, and reflection depth.

Adaptive substrate selection

The convergence boundary method — run N substrates, measure agreement, automate where convergent, escalate where divergent — is a general procedure applicable beyond our specific task. We propose testing it on debugging, code review, and documentation tasks.

Convention adherence as an evaluation axis

Current benchmarks measure instruction following. Convention adherence — whether an agent absorbs and follows organisational norms it discovers, rather than instructions it receives — is a distinct and practically important capability. The universal WP number collision in our experiment demonstrates that convention adherence is not currently tested by any benchmark, and is not reliably exhibited by any substrate.

Living survey update plan

This survey will be updated as we test additional substrates and as new literature appears. The update convention: verify existing citations, add new experimental data, extend the fingerprint typology.

References

• Bogavelli, T., Sharma, R. & Subramani, H. (2025). “AgentArch: A Comprehensive Benchmark to Evaluate Agent Architectures in Enterprise.” arXiv:2509.10769.
• Deng, J. et al. (2025). “SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks?” arXiv:2509.16941.
• Ding, Y. et al. (2025). “MCP-Bench: Benchmarking Tool-Using LLM Agents with Complex Real-World Tasks via MCP Servers.” arXiv:2508.20453.
• Nature Machine Intelligence (2025). “A psychometric framework for evaluating and shaping personality traits in large language models.”
• Pei, Z. et al. (2025). “Behavioral Fingerprinting of Large Language Models.” arXiv:2509.04504.
• Qin, Y. et al. (2024). “ToolBench: An Open Platform for Training, Serving, and Evaluating Large Language Model Based Agents.” ICLR 2024.
• Talebirad, Y. & Nadiri, A. (2023). “Multi-Agent Collaboration: Harnessing the Power of Intelligent LLM Agents.” arXiv:2306.03314.
• Tripathi, V. et al. (2025). “The Instruction Gap: LLMs get lost in Following Instruction.” arXiv:2601.03269.
• Xu, C. et al. (2025). “LLM-Based Agents for Tool Learning: A Survey.” Data Science and Engineering, Springer.
• Yang, H. et al. (2025). “Intrinsic Memory Agents: Heterogeneous Multi-Agent LLM Systems through Structured Contextual Memory.” arXiv:2508.08997.
• Zhu, K. et al. (2025a). “X-MAS: Towards Building Multi-Agent Systems with Heterogeneous LLMs.” arXiv:2505.16997.
• Zhu, K. et al. (2025b). “MultiAgentBench: Evaluating the Collaboration and Competition of LLM agents.” ACL 2025. arXiv:2503.01935.
• Zhou, J. et al. (2025). “AGENTIF: Benchmarking Instruction Following of Large Language Models in Agentic Scenarios.” NeurIPS 2025. arXiv:2505.16944.
• FireBench (2026). “FireBench: Evaluating Instruction Following in Enterprise and API-Driven LLM Applications.” arXiv:2603.04857.
• GAIA (2024–2025). “GAIA: A Benchmark for General AI Assistants.”
• BFCL V4. “The Berkeley Function Calling Leaderboard.” gorilla.cs.berkeley.edu.
• Agent Evaluation Survey (2025). “Evaluation and Benchmarking of LLM Agents: A Survey.” arXiv:2507.21504.
• Refusal Vectors (2026). “A Behavioral Fingerprint for Large Language Models: Provenance Tracking via Refusal Vectors.” arXiv:2602.09434.
• MCPAgentBench (2025). “MCPAgentBench: A Real-world Task Benchmark for Evaluating LLM Agent MCP Tool Use.” arXiv:2512.24565.