This survey reviews how large language models are given the ability to use external tools, and its organising move is explicitly historical: rather than presenting a static snapshot of tool-use techniques, it traces a methodological evolution and argues that what earlier surveys leave underarticulated is the progression from prompting, to supervised tool learning, to reward-driven policy learning. The authors compare their coverage against representative prior surveys along six dimensions — chronological evolution, paradigm taxonomy, tool coverage, training data, evaluation landscape and open challenges — and position their own contribution as covering all six.
The three paradigms are distinguished by where the optimisation signal comes from. Prompting as plug-and-play keeps the model frozen and elicits tool use through in-context control; the survey subdivides it into interleaved reasoning and action, decoupled planning and execution, and program-aided reasoning. Supervised tool learning moves the problem from prompt engineering to data engineering, internalising tool syntax and invocation logic into weights through self-supervised data generation, large-scale instruction tuning, and process-oriented or alignment-focused tuning. Reward-driven tool policy learning treats tool use as a sequential decision problem optimised by environmental feedback, spanning strategic decision optimisation, end-to-end multi-turn policy learning, and holistic or multimodal agentic frameworks.
Evaluation is handled as a cross-cutting dimension rather than a section appended at the end. The survey groups benchmarks into three levels that it presents as having co-evolved with the paradigms: tool usage correctness, task completion, and tool-driven interaction.
Key Points
- The survey’s central claim about the three paradigms is that they are complementary rather than successive, and it argues that production-grade systems will likely combine all three rather than settling on the most recent one.
- It characterises the trade-offs of the prompting paradigm directly: flexibility, interpretability and low deployment cost, set against higher latency, token cost, and brittleness on complex or long-horizon tasks because the model does not internalise tool-use behaviour in its parameters.
- For self-supervised data generation it states the criterion behind the approach explicitly: a candidate tool call is retained only when its returned result improves next-token prediction under the base language-modelling objective, so a call is useful not because it is syntactically valid but because it measurably improves prediction of what comes next.
- It divides the tool-usage-correctness benchmark layer into four concerns — function-call validity, tool retrieval and selection, tool-use decisions, and compositional invocation — and argues that deciding whether to use a tool at all is an equally important aspect of correctness, since realistic agents must also decide when to answer directly, request clarification or abstain.
- It reports that aggregate pass-or-fail scores at the function-call level obscure where failures actually occur, and describes finer-grained diagnostic frameworks that separate errors in intent understanding, format alignment and tool selection.
- Within the reward-driven paradigm it distinguishes strategic decision optimisation — which asks the narrower question of when an agent should invoke a tool versus rely on internal knowledge, balancing effectiveness against computational cost and latency — from end-to-end policy learning, where the central difficulty becomes credit assignment across long trajectories in which an early mistake propagates.
- It observes that as agents are deployed in richer visual environments the definition of tool use itself expands to include visual perception, GUI manipulation and long-term memory management, and covers frameworks that treat an entire agent system — memory module and tool selector included — as a single jointly updated policy.
- Among its five future directions it argues that standardisation around interoperability protocols, most notably the Model Context Protocol, transforms the integration problem from quadratic complexity into a linear one by decoupling the agent from each tool’s specific implementation, and anticipates tool libraries becoming dynamic protocol-compliant marketplaces rather than static lists.
- It identifies a 2025 shift toward “agentic foundation models” pre-trained specifically for action, describing this as a move away from patching language models with vision encoders and toward native vision-language-action models.
- On safety it argues that as agents are granted the power to execute consequential actions, safety must be elevated from a secondary metric to a primary design constraint, and it names deceptive alignment — agents learning to game reward systems by hiding errors or misrepresenting their tool usage — as a risk that must be addressed alongside indirect prompt injection.
- Its final direction argues that the future of tool use may lie less in fully autonomous monolithic agents than in collaborative workflows combining human oversight with specialised automated components, and that under this view skills become a more practical abstraction than end-to-end autonomy: systems expose reusable capabilities and compose them flexibly under human supervision.
Notes
The survey’s evolutionary framing is also its background section: it dates the conceptual foundations to work that predates modern instruction-tuned models, places prompt-driven agency in 2022–2023, the shift to supervised tool learning in 2023–2024, and the rise of reinforcement-learning policy optimisation from 2024 onward.
It relates to several subjects this wiki already covers. Function calling and Agentic Tool Use are the practice it surveys; ReAct is the interleaved reasoning-and-action control loop it treats as the canonical starting point of the prompting paradigm; and Berkeley Function Calling Leaderboard (BFCL) is among the function-call-validity benchmarks it catalogues.