In-depth analysis · About 5 min read
StaGPO tests how agents learn when a simulated user’s words go missing
A Notre Dame preprint groups agent turns by task needs and information gaps. Its gains depend on the training recipe, and simulated success is not proof of good human interaction.
- Published here
- Source published
- Source checked
A University of Notre Dame preprint first submitted on October 16 studies tool agents receiving incomplete simulated replies. It proposes State-Grouped Policy Optimization, or StaGPO, and reports improvements on selected Qwen3.5 tasks while documenting important failures and experimental limits.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- On ToolWOZ with a 0.4 sentence-mask rate, the reported Qwen3.5-2B recipe reaches 76.8% success versus 72.0% for base-start GRPO and 74.0% for InfoPO.
- StaGPO’s main recipe includes a supervised cold start using 33 of the model’s own dialogues. Without it, its ToolWOZ score is 67.1%, below GRPO’s 72.0%.
What the paper calls a weak user
The study does not assign ability labels to real customers. It holds the simulator’s task, knowledge and goals fixed, then replaces selected sentences in replies with an explicit inaudible marker. The opening message is never masked, and the simulator still sees its own complete history. The main simulator is gpt-oss-120b; nominal mask rates are 0, 0.2 and 0.4.
This isolates a communication channel losing information. It does not cover every way a person can be unclear, mistaken or inconsistent. Silent omissions leave no visible marker and remain untested. The authors train non-thinking Qwen3.5-2B and 4B with LoRA on ToolWOZ booking tasks and the retail domain of tau2-bench, rather than auditing general assistant subscriptions.
Exposure to missing replies was not enough
Under the standard GRPO recipe, fixed masks, mixed masks and a curriculum did not reliably outperform strong-user training on masked users beyond run-to-run noise. A smaller loss after masking can be misleading if the policy already performs worse before masking. The comparison therefore needs both the unmasked starting level and the masked score, not simply the size of the decline.
Prompting an agent to ask more also failed to produce a reliable success gain in the tested setup. This does not mean clarification is useless. In an audited subset of counterfactual prefixes, one targeted request recovered part of the information-loss cost. The paper corrects an earlier labeling problem in that probe; its useful lesson is to test whether a question closes a task requirement, not reward question count as a goal.
State is used for grouping, not as a truth oracle
Before an action, StaGPO writes a hidden structured state covering the task’s needs, missing user information and recent tool outcomes. A domain-specific key groups turns from different rollouts of the same task that reach comparable situations. The key uses known slots and outcome types rather than requiring identical full conversation text or slot values.
It then adds a step-level advantage based on discounted outcomes relative to that group to GRPO’s episode-level credit. Unanchored turns retain ordinary episode credit. The state is not supervised as an infallible representation: consistent keys can be useful even when some content is wrong. This explains why simply asking a model to print a state is a different intervention from changing how learning rewards actions in comparable situations.
The headline requires the cold-start condition
At mask 0.4 on ToolWOZ, the reported 2B recipe scores 76.8%, 4.8 percentage points above base-start GRPO at 72.0% and 2.8 above InfoPO at 74.0%. These are task-success percentages and percentage-point differences, not relative efficiency or user-satisfaction gains. The recipe includes a supervised start from 33 successful self-generated dialogues, without a stronger teacher model.
A same-start ablation helps separate the mechanism: writing states without grouping scores 60.1%, while adding grouping reaches 76.8%. Conversely, StaGPO without the cold start reaches 67.1%, below base-start GRPO’s 72.0%. The authors do not establish that the algorithm helps independently of initialization. On tau2-retail, the 4B margin over the next-best hybrid baseline is only one point, within evaluation noise.
What remains before applying this to people
The StaGPO ToolWOZ study uses 30 validation and 70 test tasks, two evaluations per test task, with selected 2B configurations repeated across three training runs. Many 4B configurations have only one training run. Task-paired errors describe task and simulator variation, not automatically uncertainty across training runs. Transfer to another masking rate is still transfer within this simulated protocol.
The authors also warn that final-state task rewards do not capture every interaction requirement, such as confirmation before acting. AISubHub’s proposed acceptance test would therefore include silent missing fields, conflicting details, explicit consent and costly mistakes alongside completion. That is a validation proposal, not a feature already proved by StaGPO. Human users, other simulators and automatically learned grouping remain open; no new API price or subscription right follows from these results.
Why it matters
AISubHub analysis: incomplete-input testing should distinguish missing information, state tracking, useful actions and final task success. More questions or a longer written state alone are not established substitutes for better training credit.
Limits and availability
We read the complete 22-page v1 HTML and appendices and checked submission history. This is an under-review, author-reported simulation study, not a human trial or our reproduction. One family, two sizes and domain-specific grouping limit generalization.
Sources and evidence
Source published: · Source checked:
Further reading
Source checked:
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.