Searcharxiv⌕ Search

arXiv · 2611.21648

Can an LLM dream? Testing the overfitted brain hypothesis on an agent with text memory

Abstract

Hoel (2021) proposed that dreams distort replayed experience so that the brain does not overfit to recent events. We test this hypothesis on an LLM agent (Claude Opus 5.5) that stores what it learns in a text memory and rewrites that memory each night from the day's conversation transcript. The agent goes through ten simulated days of conversation built from LongMemEval. We corrupt 5, 10, 20 or 40 percent of the transcript's turns before each nightly rewrite in one of two ways. Detail corruption changes one or two details in a turn to plausible wrong ones, or replaces the turn with one by the same speaker from an earlier day. Random-text corruption replaces half of the selected turns with a passage of Moby-Dick and inserts Moby-Dick words into the other half. A size limit on the memory, rising from 19,500 to 117,000 tokens over ten nights, set from human free-recall data, forces it to forget. We learned that corruption did not help the agent apply what it learned in earlier conversations to new tasks in any of 30 runs (ten conditions, three different conversation histories). Replacing details in 40 percent of turns lowered recall from 76 to 60 percent. Random text had no consistent effect on recall, with a paired difference of 0 $\pm$ 2 points at 40 percent, because the model recognized it as corrupted and left it out of memory. Replaced details were stored as fact even when another turn held the correct value, and the agent answered 40 percent of those questions correctly, compared with 75 percent without corruption. These results suggest that a pretrained model can separate corrupted text from real events, so a fair test needs a learner that cannot make this separation as easily, such as a model fine-tuned on the corrupted days or an agent whose days take place in an invented world.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Eric Zou. 2026-10-14. Can an LLM dream? Testing the overfitted brain hypothesis on an agent with text memory. https://searcharxiv.com/abs/2611.21648

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Response Attack: Exploiting Contextual Priming to Jailbreak Large Language Models

Contextual priming, where earlier stimuli covertly bias later judgments, offers an unexplored attack surface for large language models (LLMs). We uncover a contextual priming vulnerability in which the previous response in the dialogue can steer its subsequent behavior toward policy-violating content. While existing jailbreak attacks largely rely on single-turn or multi-turn prompt manipulations, or inject static in-context examples, these methods suffer from limited effectiveness, inefficiency, or semantic drift. We introduce Response Attack (RA), a novel framework that strategically leverages intermediate, mildly harmful responses as contextual primers within a dialogue. By reformulating harmful queries and injecting these intermediate responses before issuing a targeted trigger prompt, RA exploits a previously overlooked vulnerability in LLMs. Extensive experiments across eight state-of-the-art LLMs show that RA consistently achieves significantly higher attack success rates than nine leading jailbreak baselines. Our results demonstrate that the success of RA is directly attributable to the strategic use of intermediate responses, which induce models to generate more explicit and relevant harmful content while maintaining stealth, efficiency, and fidelity to the original query. The code and data are available at https://github.com/Dtc7w3PQ/Response-Attack.

cs.CL↗

Stress-Testing Emotional Support Models: Moving from Homogeneous to Diverse Help Seekers

As emotional support chatbots have recently gained significant traction across both research and industry, a common evaluation strategy has emerged: use help-seeker simulators to interact with supporter chatbots. However, current simulators suffer from two critical limitations: (1) they fail to capture the behavioral diversity of real-world seekers, often portraying them as overly cooperative, and (2) they lack the controllability required to simulate specific seeker profiles. To address these challenges, we present a controllable seeker simulator driven by nine psychological and linguistic features that underpin seeker behavior. Using authentic Reddit conversations, we train our model via a Mixture-of-Experts (MoE) architecture, which effectively differentiates diverse seeker behaviors into specialized parameter subspaces, thereby enhancing fine-grained controllability. Our simulator achieves superior profile adherence and behavioral diversity compared to existing approaches. Furthermore, evaluating 7 prominent supporter models with our system uncovers previously obscured performance degradations. These findings underscore the utility of our framework in providing a more faithful and stress-tested evaluation for emotional support chatbots.

cs.CL↗

Thinking Is Not Telling: Information Disclosure in User-Service LLM Agents

User-engaged LLM agents increasingly operate in service scenarios where task success depends on coordination between the agent, the user, and a stateful environment. In such interactions, the agent often has access to rich information that the user does not observe, such as task policies, tool results, and environment states. This makes agent-user communication a central component of task completion. In this work, we study a communication-level failure mode of user-service agents: internal reasoning is counterproductive for the interactive information disclosure. We introduce information disclosure as a measurable communication mechanism, covering state corrections, available options, constraints, consequences, and proactively retrieved environment information. Motivated by recent advances in LLM reasoning, we evaluate and analyse information disclosure under a controlled setting of agents equipped with a "think" tool, which allows agents to reason internally before responding or acting. Across five models and three user-service scenarios, we first show that enforced thinking has unstable utility. We then analyse response scaling under different think efforts and find that more thinking does not reliably translate into richer information disclosure. Using turn-level counterfactual intervention, we show that injecting golden decision-relevant information significantly intervenes agents to correct action trajectories, indicating that missing disclosure can causally affect the agent's immediate downstream behaviour. We further find that failed trajectories most often lack state corrections and constraints, and that a lightweight disclosure-aware prompt effectively mitigates degradation. These results suggest that the key challenge is not simply making agents think more, but ensuring that reasoning and information are externalised as timely information disclosure.

cs.CL↗