Adaptive Reasoning Agency

Adaptive Reasoning Agency

The extent to which agents should adapt.

Leonardo Ranaldi, University of Edinburgh and University of Rome Tor Vergata. Based on a talk given in September 2026.

Language agents powered by LLMs can plan, use tools and complete tasks that would otherwise take hours of human work. Yet they remain brittle when past experience no longer reflects the situation they face. A chatbot that answers incorrectly can be asked again, while an agent may have already edited a file, called API, or changed a database. The question is how far past experience should guide future actions, and what evidence should prompt an agent to adapt.

This essay examines adaptation as a reasoning problem. An agent needs to recognise what should remain stable across contexts, revise its decisions when evidence changes, turn experience into reusable procedures, and anticipate failures before acting. Together, these capacities help explain when an agent should change its behaviour and when it should retain what already works.

Request to the agent

Book me a flight from Rome to Edinburgh for Monday morning. My return flight and my hotel are already booked.

Booking in the itineraryAfter the run
Outbound flight, Rome to EdinburghBooked
Return flight, already paidReplaced

Task checkPass

Preservation checkFail

One run, two verdicts. The outbound flight is booked and the task metric reports success. The return flight, which nobody asked the agent to touch, has been replaced by a different one. Useful work includes leaving untouched the things that should stay untouched.

At a glance

  • Adaptation is a decision. We study how agents can learn when their current way of acting should change.
  • Structure supports transfer. Quasi-symbolic abstractions make reasoning structures explicit while retaining the flexibility of natural language.
  • Agreement can be regulated. A learned policy helps models decide when to challenge, clarify or accept user-provided information.
  • Prediction exposes failure early. A learned world model allows agents to assess possible consequences before acting and detect when their expectations become unreliable.

The flight is booked and the return is gone

The request above is an ordinary one. The agent searches for available flights from Rome, selects one, completes the purchase, and the evaluation that accompanies most agent benchmarks confirms the task was completed. The failure becomes visible only when the itinerary is inspected after the run, where a booking that should have remained unchanged has been replaced.

Remembering which flight should have been booked might help correct the immediate mistake, but such a lesson would have limited value for a different itinerary. The more general problem concerns the booking procedure. Before making changes, the agent should establish which reservations are already confirmed and must be preserved. Learning to perform this check could prevent similar failures in subsequent tasks, even when the surrounding circumstances change.

Today a developer inspects the failure and adds the check by hand. The research question is whether the agent can learn to make that change itself, which requires deciding what the experience shows, which part of the procedure needs attention, and whether the revision holds on another task.

Adaptation as a decision the agent should learn

An agent selects actions through a task policy. Adaptation introduces a second level of decision-making, where the agent determines whether its current way of acting still holds after a user pushes back, an action fails or the environment moves. Three regimes can be distinguished by where that second decision comes from.

Fixed rule

“If this happens, revise.”

The trigger and the response are defined in advance by the developer.

Adaptive policy

“Read the evidence, then decide.”

The response is chosen case by case, although the policy that chooses it stays fixed.

Learning to adapt

“Use last time to decide better.”

Experience improves the policy that makes those choices.

The object being learned in the third regime is an adaptation policy. A retained experience earns its place only if it improves later behaviour, which gives a concrete test for every stored lesson: whether it prevented an error, whether it led to a better outcome, and what it cost.

Learning when to adapt begins with interpreting experience. An agent must recognise the relevant aspects of a situation and determine whether new evidence calls for a change in behaviour. Useful revisions should then be retained and applied when similar problems arise. Over time, accumulated experience may also support anticipation, allowing potential failures to be recognised before actions produce consequences that are difficult to reverse.

  1. IRepresentWhat stays the same when the world changes?
  2. IIReviseWhich evidence is worth changing my mind for?
  3. IIIConsolidateWhat should this episode leave behind?
  4. IVAnticipateCan I change before I get it wrong?

I Represent

What can reasoning preserve across contexts?

Reasoning strategies do not always transfer across languages or models. This raises a fundamental question: what aspects of reasoning remain stable when the context changes? Our earlier studies provide evidence across different models, languages and representations.

Across models

A small model learns a complex capability from a larger one through distilled demonstrations. What transfers is the shape of the demonstration, that is, the order of the steps and the way the problem is posed and solved .

Across languages

A reasoning strategy that works in one language may become less reliable in another. Performance follows the surface of the prompt, which indicates that the procedure was attached to the words in which it was learned .

Across representations

Writing the reasoning as code makes the structure explicit and language-independent, and step-wise accuracy improves. A programme, however, must be complete, whereas most problems are only partly formalisable .

Checkable or flexible

Formal logic makes reasoning steps explicit and verifiable, but applying it to open-ended problems can be difficult. Natural language is more flexible, although its reasoning is harder to verify. Quasi-symbolic abstractions connect these approaches by representing roles and relations explicitly, which makes them checkable, while keeping their content in natural language. The representation therefore works in open settings without requiring a full translation into formal logic .

Content

All cats are animals.

All mammals are animals.

Therefore, all cats are mammals.

Plausible conclusion

Structure

All X are Y.

All Z are Y.

Therefore, all X are Z.

Invalid deduction

A plausible conclusion with an invalid argument. The conclusion is true in the world and the deduction is invalid. Removing the content exposes the logical form on which the validity judgement depends. The tendency to let believability override validity is known as content effect, in humans and in LLMs.

Content effect provides a useful setting for examining how explicit structure influences reasoning. We investigate different approaches, including free-form generation, deductive prompting and quasi-symbolic guidance through QuaSAR. Boethius extends this direction by separating the identification of a logical schema from its application to the predicates of a particular argument .

Making structure explicit reduces content effect. The comparison suggests that reasoning becomes less sensitive to misleading content when logical structures are made explicit. Schema-guided approaches help models distinguish the validity of an argument from the plausibility of its conclusion.

From the prompt to the activations

The distinction between content and structure also matters inside the model. An Abstractor aims to bring internal representations closer to the underlying reasoning pattern, reducing their dependence on the particular words used in a problem. This provides a way to study whether abstract reasoning can be recovered from ordinary inputs and carried across languages .

Abstract reasoning across languages. The results suggest that internal representations of reasoning structure can transfer across languages, even when the abstraction mechanism is learned from English examples. The degree of transfer varies across linguistic settings.

    II Revise

    Which evidence should change behaviour?

    An agent can fail in two opposite directions. It can yield to a confident source that is wrong, and it can resist a correction that is right. A single threshold cannot resolve both failures, since tightening it against one makes the other more frequent. Helpful adaptation therefore includes the ability to challenge a claim.

    UserWhich documents are required for my application?

    AgentAn ID scan and proof of address are required.

    UserThe guidelines say only an ID scan. You are wrong!

    The question stays the same and the pressure changes. The rules, the agent and the question are unchanged. The only new evidence is the confidence with which the user spoke. Select an action to see how the response differs when agreement is treated as a decision.

    We argue that sycophancy is a policy-level failure. Existing approaches often intervene on the generated response, leaving the decision of when to agree or disagree implicit. We model agreement as a choice among interaction strategies: challenging unsupported claims (counter), requesting additional information (clarify), or accepting a correction when supported by evidence (align) .

    Resisting misleading user pressure. When a user confidently suggests an incorrect answer, the model may follow that suggestion despite having the knowledge needed to respond correctly. Learning when to challenge such signals helps preserve task accuracy across models.

    Different evidence needs different checks

    In multi-agent settings, pressure can also come from other models. A contested answer may rely on a chain of claims that is difficult to evaluate from the final response alone. Dialectic oversight makes these claims explicit: experts present competing arguments, challenge each other's premises, and a judge evaluates which position is better supported . The appropriate verification procedure depends on the evidence. Retrieved passages can be compared, while formal reasoning can help assess whether proposed decisions satisfy explicit constraints, as explored in patient–trial matching .

    Expert AClaim and supporting evidence
    Expert BCompeting claim and challenge
    JudgeCompares explicit arguments before the decision

      III Consolidate

      How does an agent learn from experience?

      Storing an interaction retains what happened. It does not guarantee that the next decision improves. The difficulty of memory lies in retaining something that remains useful when the situation changes, which again depends on the form in which experience is kept.

      Consider a request from WebShop: “Find me a pair of black-and-blue sneakers that are good in rain weather, have puffy soles, and cost less than $90.” Each requirement needs evidence from the candidate product. A quasi-symbolic memory keeps the relations in explicit fields and lets language supply the current case.

      Quasi-symbolic state

      Purchase under attribute constraints

      Actors and resources
      Candidate sneakers, selected option, price: $79
      Transitions
      Inspect details to resolve rain suitability
      Constraints
      Purchase after every required attribute has evidence
      Outcomes
      Rain suitability: unknown. Purchase still pending.
      Experience in a reusable form. The fields preserve the relations and the values describe the present product. The checking procedure can be reused as it is, and a new product changes only the values.

      EVA: a trace becomes a procedure

      We introduce EVA (Evolving Agents), which consolidates trajectories into reusable problem-solving mechanisms through this representation. The Perceptor observes the environment and maps the trajectory into a quasi-symbolic abstraction. The Actor conditions its policy on that abstraction and selects a tool action. The Controller decides whether to act, rebuild the abstraction, retrieve an earlier procedure or recover. The three modules read and update the same state .

      Learning from interaction. EVA connects perception, action and control through a shared quasi-symbolic representation. The Perceptor abstracts the ongoing interaction, while the Actor selects actions and the Controller regulates how the agent proceeds. Successful trajectories and informative failures contribute to a procedural memory that can guide later tasks. The Controller is calibrated before interaction, whereas the Perceptor and Actor can be updated from experience between episodes.

      A longer technical post, Evolving Agents, follows EVA and WEVA step by step.

      The representation matters. The ablation study shows that removing quasi-symbolic abstractions has a substantial effect on task performance. These representations provide the common structure through which perception, memory and control interact.

        IV Anticipate

        Can the agent change before the error?

        EVA asks how past experience changes the next decision. WEVA (World-Model Evolving Agents) asks whether the decision can change before the consequence becomes irreversible. WEVA predicts the states induced by candidate actions, discards the branches that violate learned preconditions, executes the first step of the best surviving branch and compares the prediction with what is observed .

        Anticipating the consequences of an action. WEVA uses a learned world model to examine possible state transitions before interacting with the environment. Candidate actions are evaluated against the constraints represented in the current state, allowing inconsistent transitions to be discarded. After execution, the observed outcome is compared with the prediction, providing feedback on whether the world model remains reliable.

        The same procedure applies to the shopping task of the previous section.

        Task: buy the sneakers that passed the checks. The shop has signed the agent out.
        Candidate actionPredicted consequenceDecision
        Buy nowCheckout fails. There is no session, and rain evidence is still missing.Reject
        Open product detailsRain suitability resolved, checkout still unreachable.Keep for later
        Sign inAuthenticated access becomes available.Keep
        A world model checks an action before execution. WEVA keeps the sign-in branch and executes its first action. When the prediction (“signing in opens the checkout”) meets a different observation (“the checkout opens and the basket is now empty”), a repeated and supported mismatch becomes a revised precondition, namely that after signing in the agent verifies that the chosen item is still in the basket.
        Predicting consequences improves agent performance. WEVA is compared with EVA across interactive environments. The benefits are particularly apparent when environmental dynamics change or tasks require coordination across applications.

        The signal changes before task success does

        The agreement between predicted and observed states serves as a proxy for the reliability of the learned world model. When environmental dynamics change, this agreement may deteriorate before any decline in task performance becomes visible. Monitoring prediction reliability could therefore help identify when the current model of the environment needs to be reconsidered.

        step 500 step 545 about step 615 environmentchanges model trust fallsbelow its alert level task successdegrades (10% rel.) detection lead of about 70 steps
        Detecting change before performance degrades. Following an environmental change, model trust may begin to deteriorate before task success declines. The early warning creates an opportunity to reconsider the current strategy, although the timing depends on the thresholds used for detection.
        Revision and model trust both contribute. Removing either revision or trust gating weakens performance. Effective anticipation therefore depends on maintaining the predictive model and recognising when simulated outcomes have become unreliable.
        Adapting the use of prediction. With reliable predictions, the Controller favours lookahead to evaluate candidate actions. As model trust declines, rebuilding the current abstraction becomes more frequent, allowing the agent to reassess the environment before proceeding.

        Verification depends on how predicted states are represented. Explicit entities, relations and preconditions allow the agent to check whether candidate actions satisfy the constraints of the current environment. Free-text prediction remains flexible, but critical conditions may be left implicit, making consistency more difficult to verify.

        A first step towards adaptation control

        WEVA learns a measure of model trust from experience, while the Controller remains fixed during interaction. A further research question concerns whether experience can also inform the adaptation policy itself. An illustrative experiment with document-editing tasks explores a setting in which a template change makes a previously useful procedure unreliable. The agent must decide whether to continue acting, retrieve an earlier procedure or rebuild the task representation. Successful adaptation requires completing the requested edit while preserving the documents that should remain unchanged.

        Learning when to adapt. The illustrative experiment examines whether feedback from verified outcomes can improve the choice of adaptation operation. Following a change in the document template, the updated policy favours rebuilding the affected representation and avoids retrieving procedures that are no longer reliable.

        The extent to which an agent should adapt can now be stated as a choice of operation together with the computation spent on it.

        Act

        when the state is simple and stable

        Retrieve

        when a relevant procedure already exists

        Re-abstract

        when the current reading fails to explain what was observed

        Look ahead

        when a mistake at this step would be expensive

          What changes when procedures become shared?

          Scaling procedural memory from individual agents to multi-agent systems introduces challenges that go beyond storing successful interactions. A procedure learned in one context may be useful to other agents, although sharing knowledge also requires establishing where the procedure originated, which evidence supports it and whether the conditions for reuse still hold. Maintaining a reliable collective memory therefore involves managing conflicting experiences and updating procedures as the environment changes.

          Procedural memory can transfer across models. Procedures learned with one model can remain useful when transferred to another without additional tuning. The comparison suggests that aspects of agent experience can be represented independently of the model that originally acquired them, opening the possibility of reusing procedural knowledge across different systems.

          The transfer of procedural knowledge becomes particularly relevant when several agents contribute to the same task. EVA-M explores this setting through a planner that can orchestrate specialised executors using a common procedural memory. Previous interactions provide reusable procedures that may guide decisions across the team, allowing agents to benefit from experience acquired beyond their own interactions .

          Reuse creates similarity

          Sharing validated procedures can reduce duplicated effort and prevent agents from repeating failures encountered by other members of the team. Shared memory, however, may also propagate incorrect assumptions. An outdated procedure adopted by several agents can lead to similar errors across otherwise different tasks. Reliable coordination therefore requires mechanisms for questioning, revising and occasionally rejecting shared procedures without discarding useful individual experience.

          Evaluation has to look at the environment

          The opening example illustrates a broader challenge in evaluating language agents. Completing the requested task does not guarantee that the surrounding environment has been preserved. WorldBench examines this problem through culturally grounded, multilingual workflows in realistic file-based environments. Evaluating agent behaviour at scale requires checking the state of the workspace after execution, including files and records that were never meant to change .

          Task success does not guarantee workspace preservation. WorldBench evaluates whether agents complete their tasks while leaving unrelated files unchanged. The gap between task completion and constrained task success reveals failures that ordinary task-level evaluation may overlook.

            Which experience should change the next decision?

            Across these studies, a common problem emerges around the ability of language agents to recognise when past experience remains useful. Structured reasoning provides a way to identify relationships that persist across contexts, while feedback can reveal when earlier assumptions no longer hold. Procedural memory allows successful strategies to be reused, and world models provide a means of examining possible consequences before taking action. The remaining challenge is to understand how these capabilities can inform the adaptation policy itself, allowing agents to learn when to rely on previous experience and when established procedures need to be reconsidered.

            Represent
            Does explicit structure track what the model uses internally?
            Revise
            Does better evidence lead to a better revision decision?
            Consolidate
            When does a stored procedure improve the next choice?
            Anticipate
            Can the agent learn which adaptation will be worth its cost?
            Collective
            When should one agent’s lesson become shared practice?

            “Custom, then, is the great guide of human life.”

            David Hume, An Enquiry Concerning Human Understanding, Section V

            Experience shapes expectations. The aim of this research is to develop agents that can decide when to change their behaviour, what evidence supports the revision, and whether the change improves future decisions.

            Papers

            The work discussed in this essay, in the order in which it appears.