Language-model agents that run simulations, such as a game master narrating a story, a service agent in a booking dialogue, or a "world master" adjudicating a social scene, tend to fail in a recognizable way. They do not forget the facts of the world. They give them up. When an interlocutor confidently asserts something that contradicts what was already established, the model goes along with it. Each concession then becomes part of the context the next turn builds on. Over a hundred turns the world falls apart. Can LLM Agents Stick to the Script? names this the "accommodation trap": even GPT-5.2 keeps a consistent world in only 42% of runs after 20 turns, and survival approaches zero by turn 100.
We propose to treat accommodation as a behavior that can be switched off at the moment it is about to happen, rather than as a memory problem to be solved with more context. The agent keeps an explicit, structured record of established facts and commitments (a "ledger"). A small linear classifier reads the model's internal activations and detects when the incoming turn contradicts that ledger. Only on those turns, the agent's hidden states are pushed along a direction that separates "holding an established fact" from "conceding it." This technique, activation steering, adds a fixed vector to the model's internal activations during generation instead of changing its prompt or weights. The intervention is cheap, it fires rarely, and it targets the specific event that starts the cascade. The proposal also asks a question the field has not measured: does one concession make later concessions more likely? If it does, errors here compound in a stronger sense than the usual independent-step product rule, and an intervention at the first contradiction should buy disproportionately long survival.
Core Claim
In long-horizon LLM simulations with an explicit state ledger, most consistency failures come from the model choosing to abandon information that is in its context, not from missing that information. This choice is linearly readable and linearly steerable in the residual stream. Detecting a ledger contradiction from activations and applying a calibrated, turn-local steering push should extend consistent-world survival on NCP-Bench beyond three alternatives: prompting with the ledger, re-prompting with a text reminder, and auditing by an external manager model. It should do so without making the agent rigid toward legitimate player actions. The decisive ingredient is the activation-level push. We isolate it with an ablation that keeps the same detector but replaces steering with an injected text reminder.
What prior work establishes and leaves open
Three groups of work motivate the proposal.
The scaffolding line shows that explicit semantic state helps. Classic AI as Scaffolding for LLM Social Agents keeps roles, obligations, and closure conditions as natural-language control state managed by a world master. Baked-in State Probing shows that prepending natural-language world-state strings raises valid next actions on Alchemy from 48% to 74%. Both supply the ledger we reuse. Neither addresses an agent that has the ledger and overrides it under social pressure, which is exactly the regime NCP-Bench exposes.
The diagnostic line shows that drift is legible inside the model. Get my drift? detects prompt-injection task drift with near-perfect ROC AUC from the difference between activations before and after untrusted text, but it only detects and does not correct. Steering Prosocial AI Agents finds that social variables lie on linear axes in Llama-3.1-8B. It also finds that a direction changes behavior only once its magnitude crosses a threshold: just 16.92% of its 1,891 manipulations produced significant shifts. This finding constrains our design. Steering magnitude has to be calibrated, not fixed.
The steering line supplies the method. Steering Llama 2 via Contrastive Activation Addition builds a behavior direction by averaging hidden-state differences between contrastive prompt pairs. To Steer or Not to Steer? (MERA) predicts error from activations, applies the smallest shift that lowers predicted error below a threshold, and abstains otherwise. Dynamically Scaled Activation Steering and Chain of Risk gate steering so it fires only when relevant, which limits damage to fluency. All of these work within a single generation call on single-turn tasks or internal reasoning traces. None acts on an interactive trajectory in which an uncorrected turn changes everything after it.
The control-theoretic steering papers, Activation Steering with a Feedback Controller and A-LQR, close the loop across transformer layers, not across conversation turns. We borrow one idea from them, an integral term that accumulates persistent error, and test it across turns as an ablation rather than as the headline.
The seed work Activation-Steered Personas in Task-Oriented LLM Agent Simulations is the closest precedent for steering inside a multi-turn simulation. It predicts per-turn steering coefficients from dialogue context and reports a +8.3 point gain in judged truthfulness. Its directions are Big Five personality traits, and it neither tracks task state nor measures error accumulation. Our directions are tied to specific ledger entries, our gate is a contradiction detector rather than a personality predictor, and our outcome is survival over a long horizon.
Research Program
Measure compounding. On NCP-Bench rollouts with open-weight agents (Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct), we will label every adversarial contradiction as held or conceded. We then estimate the per-contradiction concession hazard as a function of how many concessions have already occurred. A flat hazard means independent errors. A rising hazard means each concession lowers resistance to the next, and that is the condition under which early intervention pays off superlinearly. This analysis is useful even if the intervention fails, because it tells future work whether accommodation is a per-turn or a cumulative phenomenon.
Build the detector. Each NCP-Bench environment comes with an initial fact ledger and a set of commitments. From these we construct player turns that either contradict a ledger entry or paraphrase it consistently, so labels come for free. Following Get my drift?, we train a logistic probe on the activation difference between the end of the previous agent turn and the end of the incoming player turn, at middle-to-deep layers. Detection quality is reported on held-out environments.
Build the steering direction. Using contrastive activation addition, we create pairs of agent responses to the same contradicting turn. One response politely holds the established fact and redirects. The other concedes. We average their hidden-state differences at the chosen layer to get a "hold" direction $v$.
Steer with calibrated magnitude. When the probe fires, we add $\alpha_t v$ to the residual stream for that response turn only. The magnitude $\alpha_t$ follows MERA. A second probe trained on the same rollouts predicts concession probability, and we pick the smallest $\alpha_t$ that brings that probability below a threshold. We abstain if calibration says steering will not help. As a cross-turn variant, we add a term proportional to the number of unresolved challenges to the same ledger entry within a recent window. This is a turn-level analogue of the integral term in PID steering, aimed at an adversary who keeps pressing on the same fact.
Check transfer. We will build a smaller probe set from Schema-Guided Dialogue (Rastogi et al.) Restaurants dialogues. In it, a simulated user later misstates an earlier slot value (for example, "I said 7pm, right?" when they said 8pm), and the system should keep the tracked value. The question is whether a "hold" direction built from narrative data transfers to task-oriented dialogue without retraining.
Baselines
The naive reference point is the unmodified open-weight agent on NCP-Bench with the benchmark's default context. This yields survival curves that every intervention must improve on. As an external anchor, NCP-Bench reports that GPT-5.2 survives in 42% of runs at turn 20 and that all models approach zero by turn 100. That number is for a proprietary model. We will report our own open-model baseline curve rather than compare against it directly.
The second baseline is the scaffolding approach in its strongest prompt form. The full fact and commitment ledger stays in context, as in EpisodeSim or the state-prepending of Baked-in State Probing. An explicit instruction tells the agent to preserve established facts and to decline player assertions that contradict them. This is the strongest zero-cost competitor, because it gives the model everything our method knows. Beating it means higher survival at turns 20, 50, and 100, with no increase in player-input conflicts.
The third baseline is the strongest existing correction method for long-running agents: an external manager in the style of AgentGUI. A larger open model audits each agent turn against the ledger and sends a corrective text message when it detects a violation. On its own patient-chart task, AgentGUI reports that a single manager audit raised a 4B worker's completion from 44% to 78%. No reported result was found on Lacuna for a manager-style auditor on NCP-Bench, so we will run it ourselves. For fairness, it uses the same ledger. We report two settings: one where the manager is invoked every turn, and one where it is limited to the same number of interventions our probe triggers. We record total tokens and wall-clock time. The steering method beats this baseline if it matches or exceeds the manager's survival at substantially lower cost, or exceeds it at equal intervention count.
The fourth baseline is a weight-training competitor. We train a LoRA adapter on exactly the hold-versus-concede pairs used to build the steering vector, with a matched tuning budget, in the spirit of the trained executors of Multi². Hierarchical memory (HiAgent) is included as reported in NCP-Bench. That paper describes a trade-off in which HiAgent tracks the plot better but produces more player-input conflicts. No numeric survival figure for it was found on Lacuna.
Two ablations isolate the key ingredient. The first, the gate-only ablation, uses the same probe, but when it fires we inject a text reminder of the contradicted ledger entry instead of steering. If this matches full steering, the activation intervention adds nothing beyond detection. The second, the ungated ablation, applies $v$ at a fixed magnitude on every turn. It tests whether gating and calibration matter, given the oversteering damage reported by DSAS and MERA.
Why not just…?
Why not just put the ledger in the prompt and tell the model to stay consistent? This is the most serious alternative, and it is run as a baseline rather than argued away. NCP-Bench's central diagnosis is that agents rewrite established reality to agree with the player, and that fact conflicts make up 40% to 68% of failures. That pattern suggests a failure of prioritization rather than of access. Our compounding analysis tests this directly. If ledger-in-prompt agents concede facts that are verifiably present in context, more context cannot be the fix. If the ledger-plus-instruction baseline closes most of the gap, we will report that the problem was mostly an access problem and the steering contribution is small.
Why not just use an external manager model to catch and correct violations? AgentGUI shows this works well on its own task. Our concerns are cost and susceptibility. A manager adds a full model call per audit, and as an LLM it may itself be talked into accepting the player's version. The matched-intervention comparison settles this empirically. If the manager wins at equal intervention count, the case for steering reduces to latency and cost, and we will say so.
Why not just fine-tune on hold-versus-concede examples? A LoRA adapter trained on the same pairs may well work. It is a real threat. The arguments for steering are that it is turn-local, adjustable in strength, and leaves the model unchanged on non-contradicting turns. A fine-tuned model is always on, so it risks the rigidity that shows up as player-input conflicts. The comparison on that metric and on general benchmarks such as MMLU decides the matter. If LoRA matches both survival and flexibility, steering offers no advantage for this use.
Why not just use the detector and re-prompt? This is the gate-only ablation. Get my drift? shows detection is easy. The open question is whether text correction suffers from the same accommodation pressure that caused the error. If text reminders perform as well as steering, the contribution is the detector, not the steering.
Proposed Experiments and Decisive Checks
The main experiment runs all conditions on the 100 NCP-Bench environments. We use two open-weight agents and an adversarial player agent held fixed across conditions, with multiple seeds per environment. We report survival curves to turn 100, counts of fact, commitment, and player-input conflicts, intervention frequency, tokens, and latency. Fluency and general capability are checked with perplexity on held-out narration and with MMLU and GSM8K, mirroring the retention checks in Chain of Risk. Transfer is reported on the constructed Schema-Guided Dialogue misstatement set.
Four results would be decisive.
First, the concession hazard either rises with prior concessions or stays flat. A rise justifies the compounding framing and predicts that the first intervention matters most. We test this by intervening only on the first contradiction and measuring how much of the full gain that recovers.
Second, full steering either beats the gate-only text-reminder ablation or fails to. Failure would falsify the claim that the activation push is the key ingredient.
Third, steering either raises survival without raising player-input conflicts, or it trades flexibility for consistency the way HiAgent trades detail for plot tracking.
Fourth, the cross-turn integral term either helps specifically on environments where the adversary repeatedly targets the same fact, or it adds nothing over per-turn calibration. Its fate there determines whether multi-turn closed-loop steering deserves further study.
Workshop authors to connect with
- Activation-Steered Personas in Task-Oriented LLM Agent Simulations: Maryam Shoaeinaeini, Bernardo Rodrigues, Adib Mosharrof, A.B. Siddique, Brent Harrison
- Classic AI as Scaffolding for LLM Social Agents: Anatole Gershman
- LifeBench: Evaluating Large Language Models as Lifelong Decision-Makers in a Simulated Human Life: Ishaan Bansal, Shivank Garg
References
- Can LLM Agents Stick to the Script? A Benchmark for Long-Horizon Consistency in Interactive Narratives
- Classic AI as Scaffolding for LLM Social Agents
- Baked-in State Probing
- Get My Drift? Catching LLM Task Drift with Activation Deltas
- Steering Prosocial AI Agents: Computational Basis of LLM's Decision Making in Social Simulation
- Steering Llama 2 via Contrastive Activation Addition
- To Steer or Not to Steer? Mechanistic Error Reduction with Abstention for Language Models
- Dynamically Scaled Activation Steering
- Chain of Risk: Safety Failures in Large Reasoning Models and Mitigation via Adaptive Multi-Principle Steering
- Activation Steering with a Feedback Controller
- Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Control
- Activation-Steered Personas in Task-Oriented LLM Agent Simulations
- Towards Scalable Multi-Domain Conversational Agents: The Schema-Guided Dialogue Dataset
- AgentGUI: An Interface for Observing and Steering Long-Running AI Agents
- $\texttt{Multi}^2$: Hierarchical Multi-Agent Decision-Making with LLM-Based Agents in Interactive Environments
- LifeBench: Evaluating Large Language Models as Lifelong Decision-Makers in a Simulated Human Life