Natural Language Tool Selection (NLT)¶

Describe tools in prose and parse the model's plain-language choice instead of forcing JSON tool calls — a large accuracy win on weaker models.

Natural Language Tools (NLT) present the available tools as a prose list and ask the model to pick in plain language — for example a YES/NO per tool — then a regex extracts the selection. Structured tool calling instead forces the model to emit a valid JSON function call. A 14-model replication study found NLT reached 62.3% tool-calling accuracy versus 47.4% for structured calling, a 14.9 percentage-point gain, and cut critical errors by 93% (51 versus 755) across 8,560 trials (Somma et al., 2026).

When to reach for it¶

The gain is not uniform — it tracks model capability, so lead with the condition. Choose NLT when your agent runs on models that are weak at structured output:

Small or open-weight models. Mistral 7B gained 39.4pp; it and Qwen3-VL "collapsed completely with structured outputs: 0% accuracy," mostly from failing to emit valid JSON (Somma et al., 2026).
Reasoning models. DeepSeek-R1 gained 24.0pp (55.0% versus 31.0%) (Somma et al., 2026).
Models without strong native tool calling. Claude Sonnet 4 gained 43.1pp (61.9% versus 18.8%) in this single-turn selection setting (Somma et al., 2026).

NLT also cut token usage 25.2% and won on 11 of the 14 models tested (Somma et al., 2026).

Why it works¶

Forcing schema compliance draws on a different part of the model's learned distribution than the task itself. The authors attribute the gap to cognitive-resource competition: JSON generation patterns are "primarily acquired from coding corpora, whereas tasks such as customer service or mental wellness originate from different domains," so producing a schema engages deliberate format compliance "at the expense of System 1's intuitive task understanding," diverting capacity from the actual decision (Somma et al., 2026). The effect is corroborated independently: strict JSON-mode generation degrades accuracy on reasoning-heavy tasks compared with free-form generation converted to structure afterward (Tam et al., 2024), and structured-output constraints measurably suppress correct tool calling in open-weight models (Constraint Tax, 2026).

When this backfires¶

Frontier models optimized for structured calling. GPT-5 gained only 1.6pp (both near 81%) and Gemini 2.5 Pro regressed 33.7pp under NLT (82.1% structured versus 48.3%) — the prose-parsing overhead buys nothing and can hurt (Somma et al., 2026). The study notes reinforcement-learning optimization is narrowing the NLT advantage on these models over time.
Parameterized or multi-turn tool calls. The study validates only single-turn, parameterless tool selection across two domains. A YES/NO regex parse does not extract typed arguments the way a JSON schema does, so NLT does not obviously carry to tools that take structured inputs (Somma et al., 2026).
Systems needing a guaranteed machine-readable contract. Replacing schema-validated JSON with regex-parsed prose adds a parse-failure surface that structured calling eliminates by construction. See the trade-off in Structured Output Constraints.

Example¶

The NLT prompt lists each tool as one prose line asking for a per-tool decision, wraps it in a reasoning slot and an end marker, then a regex reads the answers:

Thinking: <the model's reasoning about the request>
Website information – YES/NO
<one line per available tool> – YES/NO
Assessment finished

The model replies in prose (Website information – YES) and the harness extracts the selection by exact match — no JSON, no schema validation, no failed-parse retries (Somma et al., 2026). The structured equivalent passes OpenAI-compatible function schemas and requires the model to return a valid tool_calls JSON object, which the weakest models fail to produce at all.

Key Takeaways¶

NLT lifts single-turn tool-selection accuracy 14.9pp on average and cuts critical errors 93% across 14 models (Somma et al., 2026).
The gain is capability-dependent — large on small, reasoning, and non-native-tool-calling models; near-zero or negative on frontier models optimized for structured calling.
The mechanism is distributional: schema formatting pulls from coding-corpus patterns and competes with task reasoning.
Keep structured calling for frontier models, parameterized tools, and multi-turn flows that need typed arguments or a machine-guaranteed contract.
Decide per model, not once for the whole system — pair it with Per-Model Harness Tuning.