acceptodds
Under review as a conference paper at ICLR 2027

DuoGate: Action and Context Control for Reliable Language Agents

Abstract

LLM agents can report success without executing the actions a task requires. When runtime checks correct such failures, their explanations provide learning signals beyond the final task outcome. Using this feedback effectively requires deciding when an interaction warrants learning and which lessons to retain. We introduce DuoGate, a framework that connects action control with selective context adaptation through execution feedback. Its action branch checks selected proposals against available evidence, directs recovery toward detected gaps, and preserves correction feedback for learning. Its context branch uses outcomes and execution signals to select learning updates. For these updates, a reflection and curation workflow derives candidate guidance from the interaction and correction reasons. Separate admission checks screen this guidance before it enters a playbook for later prompts, so an intervention need not become a retained lesson. We evaluate DuoGate on AppWorld, -Bench, and Terminal-Bench 2, finding higher mean completion than the adaptive context baseline across all three benchmarks. With MiniMax M2.7, DuoGate raises AppWorld task completion from the baseline's 60.7% to 69.2% and scenario completion from 41.7% to 54.2%. These results demonstrate the complementary benefits of action control and context adaptation, highlighting correction feedback as an effective signal for learning from execution. Code: https://anonymous.4open.science/r/DuoGate-56CF/.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LLM agents, Context Adaptation, Action Control, Agent Memory

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ggZHOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DuoGate combines action verification with selective learning from execution feedback. It checks proposed actions against available evidence and uses detected failures to guide recovery. It also uses task outcomes and correction feedback to decide which experiences to retain. Before any resulting guidance is added to the playbook for future prompts, it is screened to reduce the risk of carrying mistakes forward. The evaluation covers AppWorld, τ²-bench, and Terminal-Bench, with ablations on action control and context updates.

Critical strengths
  1. The paper addresses a practical agent failure: claiming task completion without executing the required actions. It combines runtime verification with reusable guidance, and shows consistent gains across multiple models and benchmarks.

  2. The ablations examine both learning-opportunity selection and guidance screening. The explanation studies offer useful insights into what drives the gains, while the reporting acknowledges limitations in statistical evidence and trajectory analysis.

Critical weaknesses
  1. Incremental novelty is unclear. DuoGate builds on ACE’s reflection and curation, but lacks a matched comparison isolating the gains from its specific coupling of action control and learning. Its advance over ACE[1] and related work such as MPR[2] remains insufficiently established. MPR uses frozen memory at inference, so the concern is the additional value of DuoGate’s online coupling rather than equivalence between the methods.

  2. Limited evidence for diagnostic gains and generalization. Original feedback yields only marginal gains over generic verification (0.40 pp in task completion and 0.59 pp in scenario completion; Table 13). Fixed task ordering and related AppWorld variants also leave order sensitivity and broader transfer untested.

[1] Zhang et al. Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models. Section 3. https://arxiv.org/html/2510.04618v1 [2] Wu and Qu. Meta-Policy Reflexion: Reusable Reflective Memory and Rule Admissibility for Resource-Efficient LLM Agents. Sections 3.2–3.5. https://arxiv.org/html/2509.03990v1

Questions for authors

What is the incremental benefit of DuoGate’s coupling over the same action controller with standard ACE learning, and how does it advance beyond ACE [1] and MPR [2]?

Given the small gap between Original and Generic feedback, what evidence supports the added value of detailed correction diagnoses and the generalizability of learned guidance?

Reviewer_ZqnkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DuoGate checks agent actions at runtime and uses selected interactions to update a playbook. It controls both when to learn and which guidance to keep. Experiments cover AppWorld, tau2-Bench, and Terminal-Bench 2.

Critical strengths

The action and context ablations test whether intervention actually helps later learning. Results across four AppWorld backbones support the combined design. I also like that the controller distinguishes permission to act, successful execution, and evidence for claiming success. The explanation-substitution experiment is especially useful because it holds runtime feedback fixed and changes what reaches the learning stage. This makes it possible to examine the learning contribution separately from immediate recovery.

Critical weaknesses

The cost comparison leaves out assessor calls used by the controller. It therefore does not tell us whether DuoGate costs less overall than ACE.

I am less convinced that the detailed correction explanations are necessary. In Figure 5, generic feedback is only 0.4 points behind on tasks and 0.6 on scenarios. A reminder to verify may account for much of the benefit attributed to explanation content. The generic-feedback result does not undermine the combined system, but it narrows the explanation for why it works. I would accept a more modest claim here: selective feedback helps, while the value of the detailed correction text remains uncertain.

Questions for authors

What is the total token and monetary cost with assessor calls included? Does the gap between original and generic feedback persist across task orders, with playbooks reset for each run? For cost, the relevant comparison is the complete pipeline on the same sequence of tasks.

Reviewer_9qBfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses a reliability problem with LLM-based agents, an agent may claim it finished a task but it never actually ran anything, or it goes ahead and acts on a guess.

The authors propose DuoGate, and it's built with two branches that work together. The first is an action control branch that steps in before any tool gets called. It checks the proposed action on four dimensions, whether the inputs are valid, whether the action is permitted, whether the output is grounded in evidence and whether the task is really done. If any of those checks fail, the action is suspended and the system tries to recover via filling in a parameter that was missing, asking the user what they meant, or pushing the agent to keep going.

The second is selective in-context reflection learning, and the motivation is compute cost. Because actions are already being checked as they happen, there's no need to reflect on every trajectory. Reflection only steps in when needed at runtime, e.g. interrupted action, tool-calling exception, or the agent exceeding its step limit. When it step in, the agent gets real feedback about what went wrong rather than only pass/fail signals, so it can actually get the root cause instead of just retrying.

Lastly, there's an admission gate that screens new entry before it enters a playbook for later prompts. It drops redundant, excessively long, or improperly formatted context, which keeps the playbook short and usable.

Critical strengths

The main strength of the paper is that the two branches work complementarily. Running both at once outperforms either one alone in all 13 evaluation settings, and in several cases, the combination succeeds where neither branch on its own could outperform the base agent.

The authors demonstrate that the runtime diagnostic explanations have a positive impact, giving a 6 pp gain in scenario level generalization, as they're not just single pass/fail checks. Furthermore, they also keep reflection separate from playbook admission, which avoid the prompt wasting and keep token costs down, while this doesn't sacrifice performance.

Critical weaknesses

One concern is how much of the framework is hand-built. There are serveral domain-specific rules, which makes it hard to tell how much of the gain would transfer to other settings, and how much is careful engineering for particular benchmarks.

The comparison against ACE is also uneven. The gains are real on AppWorld (+8.5 pp) and Banking, but on Terminal-Bench 2 the improvement is only 2.6 pp, and on τ2\tau^2-Bench Retail the two are basically tied.

Since the calls of LLM-as-a-judge are part of the sync action checking, the actual compute cost of running the framework is higher than what the numbers suggest.

Questions for authors
  1. What do the total token count and latency look like once the sync LLM-as-a-judge calls are included ?

  2. How much of the gain comes from careful engineering, and how much from the proposed design?

  3. Why are the gains over ACE so small on Terminal-Bench 2 and τ2\tau^2-Bench Retail, but AppWorld shows a much bigger gain?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.