Stop stuffing expert trajectories into the prompt: one fitted concept vector lifts the three-benchmark average to 75.9 as the model grows

An agent choosing actions in sequence does not need full expert trajectories in its prompt, nor a natural-language summary of experience, if expert choices at each step directly supervise a reusable scoring rule. Earlier routes either put whole trajectories into the prompt, where they compete with interaction history for limited space, or summarized experience in natural language and left the agent to translate general guidance into a specific action. Across TextWorld, ALFWorld, and ScienceWorld at Qwen3-4B, 8B, and 14B, CB-Agent reaches cross-benchmark averages of 66.6, 72.5, and 75.9, the highest among the evaluated methods, and ranks first on TextWorld and ScienceWorld at every size.

The method turns expert trajectories into decision examples and scores candidate actions with frozen language-model features, where the concept vector only adjusts the language-model prior and the language model itself is never updated. The concept is fitted by regularized maximum likelihood, only that one vector is optimized, and candidate features and prior weights are computed before fitting and held fixed. In an ALFWorld control that sets the concept term to zero, keeping the learned concept raises success rates by 10.4, 6.0, and 14.9 percentage points at 4B, 8B, and 14B, indicating the gain comes from the learned concept rather than the length-adjusted language-model prior alone.

The concept stays fixed during a task while action probabilities change with the context and candidate set, so one scoring rule can be reused across test tasks of the same task type. The ReAct prompt still carries demonstrations to guide reasoning and think-or-act routing, whereas the action scorer uses only the current task and interaction history, with no expert demonstrations. In a two-trial ALFWorld setting, combining CB-Agent with Reflexion adds 4.5, 7.5, and 5.2 percentage points at 4B, 8B, and 14B, with the concept not refitted from test feedback.

The method targets sequential action selection in text-interaction environments and is evaluated on TextWorld, ALFWorld, and ScienceWorld at Qwen3-4B, 8B, and 14B, with demonstrations fixed per task type and one concept fitted per task type per backbone and reused across test tasks of that type. Fixed-action-space tasks get candidates from the environment, while open-action-space tasks get a finite candidate set sampled by the model. It suits readers interested in experience reuse and action scoring for language agents. Next steps are examining candidate coverage and transfer to unseen task types.

The theoretical analysis of the concept-fitting objective covers fixed features and prior weights, while end-to-end task performance is assessed experimentally. In open-action-space tasks, action selection depends on the coverage and quality of generated candidates, and candidate generation, scoring, and concept estimation add computation, with comparisons assessing task performance rather than efficiency under matched computational budgets. Demonstrations are omitted from the action-scoring context but remain in ReAct reasoning and routing, so the complete agent is still subject to context-window limits. CB-Agent does not lead on ALFWorld, and the authors suggest one possibility is that better individual action choices may still leave errors that prevent task completion.