Specialized OODA, the stage-specific version of OODA-Tool, recorded the highest Task Success score at every reported model size in an arXiv version 1 preprint dated 25 August 2026. On Qwen3-Instruct backbones ranging from 0.6B to 14B parameters, its scores were 85.10 at 0.6B, 85.46 at 1.7B, 87.30 at 4B, 89.52 at 8B and 94.90 at 14B. Against Direct-LoRA, its reported Task Success margin was 6.86 points at 0.6B, 6.79 at 1.7B, 6.99 at 4B, 5.94 at 8B and 4.48 at 14B. The gap narrowed at the largest model but remained positive across the reported range.
OODA-Tool treats tool use as a closed loop: Observe, Orient, Decide and Act happen in sequence, with a central controller validating the stages and feedback starting the next cycle. In the comparison, Joint produced the typed result in one pass; Shared used four stage-wise calls with one adapter; Specialized used four stage-wise calls with a separate adapter for each stage. All variants used the same Qwen3-Instruct backbone, tokenizer and context budget.
The main evaluation used the official ToolDial split of 11,111 multi-turn tool-use sessions. It tested Qwen3-Instruct models at 0.6B, 1.7B, 4B, 8B and 14B parameters, measuring Task Success, Tool Exact and Ask–Act Accuracy. Specialized OODA led on Task Success at every scale; Tool Exact followed the same general pattern, while Ask–Act Accuracy was near saturation for structured methods.
Where the extra structure helped
The advantage was uneven across the workload. Relative score differences were largest on long-history, missing-information, state-change, active-constraint and deep-dependency tasks—cases where the agent must reconstruct or update what it knows before acting. They were much smaller for parallel calls, where action expansion and cross-call argument binding remained the main obstacles.
An ablation analysis—tests that remove one part of a system at a time—covered 1,250 turns from 600 dialogues. Removing Orient produced the largest degradation. Removing the Decide–Act split was especially harmful on missing-information and multi-tool turns, while removing stage-specific adapters had the smallest effect.
Transfer tests also favored Specialized OODA in the reported settings. At 1.7B, it was best or tied on the shared FAIL-TaLMs readiness measures and improved most MTU-Bench metrics while tying the rest. On BFCL, the clearest advantage appeared with smaller backbones; the margin decreased as models grew, and the gain on the Multi-turn track was smaller, in line with the parallel-call errors seen in the main analysis.
The cost of remembering
The method also carried a timing trade-off. At 1.7B, Specialized OODA used four sequential calls and its normalized latency—a relative timing measure—was 2.36 times Direct-LoRA's. Direct-SC@4 used the same number of model calls but still remained below the OODA variants.
The study also examined whether the information passed between stages stayed useful. Removing downstream access to the full conversation history retained most Full OODA performance, but weakened long-history, missing-information and multi-tool turns. Removing typed states was less robust, while removing Orient caused the largest losses in state reconstruction and constraint consistency, even though Ask-vs-Act accuracy stayed near the ceiling.
Argument grounding—using the identifier that belongs to the selected item—made the issue concrete. On a fixed 500-turn sample drawn from rare-error pools, Specialized OODA had the lowest premature-call rate and the lowest stale-or-ungrounded-argument rate among the compared methods. In a worked example, the user selected osm_id_001 and the downstream tool required a Wikidata identifier. OODA extracted Q123456 from the selected row, while Direct-LoRA and ReAct-LoRA copied museum_001.
On the specified turns, the State–Action Contradiction Rate, or SACR, was 7.8% for Joint, 5.6% for Shared and 3.9% for Specialized OODA. Without the Decide–Act split, it rose to 10.7%; the gap between Specialized OODA and that no-split version was 6.8 percentage points.
A targeted trade-off
Taken together, the reported pattern is strongest for state-intensive work, while parallel-call expansion and cross-call argument binding remain difficult. Typed states do not preserve every piece of useful context on long-history, missing-information and multi-tool turns. The clearest transfer advantage appeared for smaller backbones and was smaller on the Multi-turn track, while Specialized OODA's four sequential calls carried a latency cost.
Paper data and sources
Original title: From State to Action: OODA-Tool for Reliable Multi-Turn Tool Use
Authors: Rongfeng Guo, Yinxuan Huang, Yusen Wu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text