Open problems

A recent study reports that the representations a language model is poised to verbalise behave like a global workspace: their contents can be reported, held and used in silent reasoning, and they carry coherent content only in an intermediate band of layers^[23]. Measurement in interpretability has known traps. A probe can learn the task itself rather than reflect the representation, which control tasks are built to expose^[24]. A subspace intervention can change behaviour through a dormant parallel pathway, giving an illusory attribution^[25]. Before such measures are used to compare models, or to train a model to have a workspace, their repeatability and their agreement with one another need to be shown.

Editing methods are usually judged on whether the edited fact is recalled and unrelated facts are left alone. Benchmarks that test the consequences of an edit find that methods which recall the edited fact fail badly on multi-hop questions whose answers should change^[26], and fail to make consistent changes across related facts, where a simple in-context baseline scores best^[27]. Where causal tracing localises a fact does not indicate which layer is best to edit^[28]. A system that corrects facts over time needs to know how propagation and collateral damage change as the number of dependent facts grows, which an average over edits does not show.

Networks trained on tasks in sequence can lose earlier learning abruptly^[29]. In continual learning, examples learned quickly are less prone to forgetting, and choosing replay examples by learning speed reduces forgetting^[30]. Separately, language models can be trained to predict whether they know the answer to a question, though that prediction is poorly calibrated on new tasks^[31]. These lines leave a narrower question open: whether a model can predict, from a description of a coming disruption, which of its memories will fail, and whether a fixed protection budget spent on that forecast does better than one spent on simpler signals.

Long-term conversational memory remains hard. On LoCoMo, long-context models and retrieval improve results but still lag substantially behind human performance^[32]; on LongMemEval, commercial assistants and long-context models show a 30% accuracy drop across sustained interactions^[33]. Systems add structure in different ways: tiers of memory managed like an operating system, as in MemGPT^[34], or a temporally aware knowledge graph, as in Zep^[35]. Their evaluations overlap only on one benchmark, Deep Memory Retrieval, on which Zep was compared directly with MemGPT, and each reports evaluations the other did not run^[34]⁠^[35]. The open question is which layer of structure earns its cost on which failure mode, such as knowledge updates or temporal reasoning, when all are compared under one protocol with cost reported.

Ontology-driven self-training improved a model on hypernym discovery and medical question answering, and preserved more of its original knowledge than TaxoLLaMA, a direct ontology-injection method^[36]⁠^[37]. Against that, models struggle to learn new facts through unsupervised fine-tuning, and retrieval outperformed fine-tuning on knowledge-intensive tasks^[38]. Grammar-constrained decoding lets an untrained model meet a required output structure, and on some structured tasks it beats task-specific fine-tuned models^[39]. These results use different domains, tasks and baselines, so they do not settle when structure belongs in a model's weights and when it belongs beside the model.

Small models trained on carefully selected and synthetic data can do well on narrow tasks: a 1.3-billion-parameter code model trained largely on textbook-quality data reached 50.6% pass@1 on HumanEval^[40]. A position paper argues that small models suit many calls inside agentic systems, which are specialised and repetitive, and cost less to run^[41]. Neither shows that a specialist improves a system that already pairs a frontier model with good retrieval. The comparison that matters to a builder is frontier model alone, frontier model with structured retrieval, specialist alone, and frontier model with specialist, on the same held-out tasks.

Graph-based retrieval was proposed for global questions about a whole corpus, which plain retrieval handles poorly because they are a summarisation task rather than a lookup^[42]. A systematic comparison under one protocol found distinct strengths for each approach across tasks and evaluation perspectives, and gains from combining them^[43]. That comparison used widely used text benchmarks. In domains with a governed vocabulary and explicit relations, a graph can be built from the ontology instead of being extracted by a model, and whether the same boundary between the two approaches holds there is a separate question.

Interactive world models can now generate action-controllable environments learned from unlabelled video alone^[1]. A limited temporal context, however, often breaks long-term consistency: a scene revisited after a long gap no longer matches what was generated before^[2]. One recent approach, WorldMem, adds an explicit memory of past frames and states, retrieved by pose and timestamp, and reports accurate reconstruction of previously observed scenes even after large viewpoint or time gaps^[2]. WorldMem also shows, in qualitative figures, objects the agent placed still present on return; its quantitative results use image reconstruction metrics, and changes such as plant growth are shown only qualitatively^[2]. The cited work does not include a benchmark with a ground-truth record of persistent state that scores state errors on return as a function of gap length.

World models are mostly compared on how realistic their output looks. In the few models studied so far, visual realism has not tracked physical understanding or action-following^[3]⁠^[4]. Video models that predict in a learned representation space show an understanding of object permanence and shape consistency under a violation-of-expectation test, while pixel-space prediction stays closer to chance^[5]; that test was adapted from developmental psychology for machine learning in earlier work^[6]. These studies cover few models, so they do not establish the general relation between realism and correct dynamics across model families.

Real-robot testing is costly, so action-conditioned video models are being used as stand-in environments. One such system reports that policy success rates inside the world model correlate highly with real-world success rates and preserve rankings across policy versions, sizes and checkpoints^[7]. A separate line of work argues that a world model used as a test oracle is itself an unverified learned artefact, and proposes a ladder of admissibility checks drawn from safety-critical simulation practice that a model must pass before its verdicts count as evidence^[4]. Neither cited study maps where agreement breaks: which tasks, which kinds of policy failure and which distances from the training data make a verdict untrustworthy.

Probing shows what a representation encodes, but probe accuracy need not track what a model uses; removing a property from the representation and measuring the effect on behaviour is a stronger test^[8]. In vision-language-action policies, sparse directions in activation space, such as speed and direction, have been found to be causally linked to action selection^[9], and open models such as OpenVLA make this kind of work possible^[10]. Evaluation still rests largely on end-task success, which hides weaknesses in spatial reasoning and dynamic adaptation^[11]. Neither probing nor success rates asks directly which scene variables a policy keeps steady when they are pushed, or whether that predicts where it fails.

Conditioning a language model on demographic backstories can reproduce the response distributions of many human subgroups, a property called algorithmic fidelity^[12]. Agreement with published data is the usual evidence for it, but a model may have read the published data during training. Testing whether specific text was in training is unreliable: membership inference attacks barely beat random guessing in most settings across a suite of models^[13], and on published evaluation sets blind baselines that never look at the model outperform them^[14]. So agreement with published results cannot yet be taken to separate a model's estimate of a population from its recall of a published table.

Persona-prompted model answers can match the averages of a national election study while showing less variation than the real survey, with regression coefficients that often differ significantly from estimates on the human data; minor changes in prompt wording also shift the synthetic distribution^[15]. Analysis and human studies across 16 demographic identities argue that current training leads models to misportray and flatten demographic groups^[16]. Misalignment between model opinions and those of US demographic groups is substantial and persists after explicit steering towards a group^[17]. Fidelity on averages is therefore a weak certificate for any analysis that depends on spread or on relationships between variables.

Agents built from interviews predicted participants' held-out General Social Survey answers at 83% of the participants' own test-retest consistency, better than agents given demographics only^[18]. Earlier work replicated several classic experimental findings with simulated samples and also found a systematic distortion in some models^[19]. Accuracy normalised by each person's own consistency is a strong standard. In that study it was measured on one national sample and a bounded set of outcomes; the study does not establish how it holds for behaviour far from the interview's content, over longer spans of time, or for people poorly represented in training text.

Because language models learn mostly from observational text, an intervention given to a simulated user can shift that user's implied attributes, so the treatment and control arms no longer describe the same population^[20]. Instructed chatbots drift from their system prompt within eight rounds of conversation^[21]. Model responses to changes in survey wording often fail to match human response biases, and react to perturbations that leave humans unmoved^[22]. Measuring negative-control attributes that should not change can reveal drift, and eliciting targeted confounders into the persona reduces the bias in some settings^[20]. The cited work does not show whether any simulator design keeps fixed traits fixed across long, adaptive exchanges.