Agent orchestration frameworks have proliferated, collectively exceeding 290,000 GitHub stars across LangGraph, CrewAI, Google ADK, OpenAI Agents SDK, Semantic Kernel, Strands, and LlamaIndex. All follow the same pattern: an external orchestrator above the LLM, injecting instructions and routing decisions every turn. Recent work has shown this architecture is dominated for procedural tasks by simply providing the procedure in a frontier model's system prompt [Dennis et al., 2026a], at the cost of consuming the context window, requiring a frontier model for every conversation, and exposing proprietary procedures to third-party providers. Compiling the procedure into the weights of a small fine-tuned model – creating a subterranean agent – should resolve all of these concerns, and prior work (SimpleTOD, FireAct, SynTOD, WorkflowLLM, Agent Lumos) has shown the technique works. Yet developer adoption has overwhelmingly favored orchestration. We identify three perceived barriers and address each empirically across travel booking (14 nodes), Zoom support (14 nodes, product-specific knowledge), and insurance claims (55 nodes, 6 decision hubs).

Major LLM platforms deploy models in an inference-only configuration: the model serves requests but never updates per-user weights. Users must repeatedly re-teach preferences, corrections, and project context, and context-based workarounds consume context-window space and degrade under cascading compaction. We evaluate an alternative: nightly consolidation of interaction knowledge into model weights via reflection, synthesis, and Low-Rank Adaptation (LoRA) fine-tuning on a single consumer GPU. Across ten realistic software development conversations (n = 10, 1,146 test questions across three memory types), three cycles of cascading compaction retain 36.8 +/- 3.0

Parameter-efficient fine-tuning methods like LoRA have become the default for adapting large language models, succeeding across instruction following, style transfer, and factual adaptation. We show that for procedural knowledge–the ability to follow multi-step procedures with conditional branching through to terminal states–LoRA fails to match full fine-tuning at the ranks where it retains its efficiency advantage. In a systematic ablation (r = 16–128) on a procedural travel booking task (14 nodes), all LoRA configurations fail uniformly (task success <= 2.54 vs. 4.11 for full fine-tuning, all p < 0.001), with scores decreasing at higher ranks–despite maintaining 95–99

Agent orchestration frameworks – LangGraph, CrewAI, Google ADK, OpenAI Agents SDK, and others – place an external orchestrator above the LLM, tracking state and injecting routing instructions at every turn. We present a controlled comparison showing that for procedural tasks, this architecture is dominated by a simpler alternative: putting the entire procedure in the system prompt and letting the model self-orchestrate. Across three domains – travel booking (14 nodes), Zoom technical support (14 nodes), and insurance claims processing (55 nodes) – we evaluate 200 conversations per condition using LLM-as-judge scoring on five quality criteria. The in-context approach scores 4.53–5.00 on a 5-point scale while a LangGraph orchestrator using the same model scores 4.17–4.84. The orchestrated system fails on 24

This study investigates the capability of a non-reasoning large language model (GPT-4o) to generate private speech and evaluates its similarity to human private speech. We placed the model in a simulated solitary block-construction scenario via textual prompts, eliciting and classifying its self-directed utterances using an established semantic framework for categorizing private speech in children. The distribution of these categories was compared to two human benchmarks: a classic block-construction study and a more recent experiment employing a similar task setting. Analysis using scatter plots and Pearson correlation coefficients revealed a striking pattern: GPT-4o’s semantic profile showed negligible similarity to the classic benchmark (r = 0.01) but very strong similarity to the recent benchmark (r = 0.93). This discrepancy is interpreted as stemming from differences in task nature, namely goal-directed, scaffolded task versus self-determined, unscaffolded play, which exert a stronger influence on speech content than experimental subject difference between GPT-4o and children. In an exploratory serial recall study, we tasked GPT-3.5-Turbo-instruct and observed incidental private speech, indicating that the phenomenon extends across contexts. This provides an avenue for investigating LLM replication of private speech and, potentially, computational consciousness.

Mean cross-entropy is the standard validation metric for language models, but it can fail to track model quality during training. We examine this in two common scenarios. First, in Qwen2.5-1.5B SFT on synthetic fact-learning, we find that mean CE rises substantially after the initial learning phase while held-out fact-recall accuracy remains near its peak. Second, we find that in top-K distillation on TinyStories, decreasing K improves median CE while worsening mean CE; the Top-5 student attains the highest LLM-judge score and crosses below its teacher on median CE, despite having the worst mean CE. In both cases, median CE correlates much more closely with task performance than does mean CE. Analyzing how bulk and tail percentile CE move during training reveals that training reshapes the empirical per-token CE distribution. In top-K distillation, smaller K yields a distribution with more mass at both extremes, decreasing the median and increasing the mean. In Qwen SFT, the bulk saturates quickly while the tail extends in the latter half of training. In both, the task-evaluation metric appears more sensitive to the bulk than to the tail. Practically, we recommend reporting a small set of percentile CE summaries alongside the mean, and using concordance among them as a tool to keep track of distribution reshaping, as well as a low-cost diagnostic for when mean and median CE disagree on model selection.

Research in event cognition has shown that how people segment events plays an important role in how they are perceived and remembered. One potential source of information for anticipating upcoming event boundaries is the typical duration of the event. The small amount of prior work that exists suggests that the durations of daily events have normal distributions and that people have accurate prior beliefs about these durations (Griffiths Tenenbaum, 2006). This research study aims to further explore the temporal distribution of self-reported events from daily life. Forty-eight participants provided information regarding the duration of events from their daily life for 14 days. Event durations for various activity types were characterised using duration modelling with truncated normal, exponential, and gamma models. Predictions based on mean and median values, including ± 5

BackgroundMindfulness meditation has demonstrated modest benefits for mental health and well-being, although the relationship between practice dose and outcomes is unclear. Meta-analyses and randomized controlled trials have shown mixed results so far, although such results may stem from methodological issues rather than reflecting the absence of an underlying effect. Research outside structured programs suggests that long-term practice time is linked to positive outcomes, but bias due to self-selection over time may explain these results. ObjectiveThe proposed trial aims to test dose-response effects for an online mindfulness meditation course, examining outcomes and participant engagement across different practice doses. In this pragmatic randomized controlled trial, we hypothesize that larger doses of mindfulness training will yield significantly larger effects and different doses will be significantly associated with variation in participant engagement, with lower engagement evident for higher doses. MethodsAt least 688 healthy adults aged between 18 and 65 years will be randomized to join one of three 4-week online mindfulness courses with daily practices of varying lengths (ie, 10, 20, or 30 min) against a minimally active control condition (4 min). Psychological well-being will be measured using the Warwick-Edinburgh Mental Wellbeing Scale at the baseline, midintervention, and postintervention time points and at 1-month follow-up. Secondary outcomes are psychological distress, anxiety, depression, social anxiety, nonattachment, trait mindfulness, decentering, equanimity, repetitive negative thoughts, emotion regulation, attention control, and emotional reactivity. Other outcomes will be collected weekly and daily during the intervention period. The primary analysis will be undertaken following the intention-to-treat approach. We will also conduct per-protocol secondary analyses on all outcomes (ie, primary and secondary). In addition, we will systematically monitor for possible adverse experiences. ResultsThis study began screening and recruitment in May 2024. Recruitment was paused approximately 6 weeks later after a substantial number of participants were identified as being fraudulent and not meeting the eligibility criteria. Recruitment reopened in October 2024, and by the end of 2024, a total of 70 eligible participants were enrolled. Recruitment recommenced in early 2025 and will continue until the end of March 2025 or until the target sample is reached. We estimate that the results will be published by March 2026. ConclusionsThis study will contribute to the evidence base for mindfulness meditation and the question of how much practice people need to engage in to improve well-being and other psychological outcomes. International Registered Report Identifier (IRRID)DERR1-10.2196/72786

The list-length effect has been crucial in understanding the sources of forgetting in recognition memory, especially whether forgetting stems from interference generated by other items in the study list. However, there has been inconsistent evidence for the effect, and recent studies have found various confounds in the experimental design. The current study re-examined the list-length effect by controlling confounds that have been reported. We also systematically manipulated various factors such as the list length, study time, delay, and stimulus type, and tested a large sample to examine the robustness of the effect (i.e., radical randomization). Results showed evidence for a list-length effect across the conditions, and we find that the square root function best describes the list-length effect. By utilizing a computational model we also show that although the list-length effect exists, there is, in general, a greater amount of interference stemming from other sources in recognition memory (e.g., previous contexts).

Introduction: Mindfulness meditation has demonstrated modest benefits for mental health and wellbeing, although the relationship between practice dose and outcomes is unclear. Meta-analyses and randomized controlled trials have shown mixed results so far, although such results may stem from methodological issues rather than reflecting the absence of an underlying effect. Research outside structured programs suggests that long-term practice time is linked to positive outcomes, but bias due to self-selection over time may explain these results. The proposed trial aims to test dose-response effects for an online mindfulness meditation course, examining outcomes and participant engagement across different practice doses. In this pragmatic randomized controlled trial, we hypothesize that larger doses of mindfulness training will yield significantly larger effects, and different doses will be significantly associated with variation in participant engagement, with lower engagement evident for higher doses. Methods and analysis: At least 688 healthy adults aged between 18-65 years will be randomised to join one of three 4-week online mindfulness courses with daily practices of varying lengths (i.e. 10-mins, 20-mins, or 30-mins) against a minimally active control condition (3-4 mins). Psychological wellbeing will be measured using the Warwick-Edinburgh Mental Well-being Scale at baseline, mid-intervention, postintervention, and 1-month follow-up. Secondary outcomes are psychological distress, anxiety, depression, social anxiety, nonattachment, trait mindfulness, decentering, equanimity, repetitive negative thoughts, emotion regulation, attention control, and emotional reactivity. Other outcomes will be collected weekly and daily during the intervention period. The primary analysis will be undertaken following the intention-to-treat approach. We will also conduct per-protocol secondary analyses on all outcomes (i.e. primary and secondary). We will also systematically monitor for possible adverse experiences. Discussion: The study will contribute to the evidence-based for mindfulness meditation, and the question of how much practice people need to engage in to improve wellbeing and other psychological outcomes. Trial registration number: The study has been prospectively registered at ClinicalTrials.gov with the identifier NCT06378450.

Models of word meaning that exploit patterns of word usage across large text corpora to capture semantic relations, like the topic model and word2vec, condense word-by-context co-occurrence statistics to induce representations that organize words along semantically relevant dimensions (e.g., synonymy, antonymy, hyponymy, etc.). However, their reliance on latent representations leaves them vulnerable to interference, makes them slow learners, and commits to a dual-systems account of episodic and semantic memory. We show how it is possible to construct the meaning of words online during retrieval to avoid these limitations. We implement a spreading activation account of word meaning in an associative net, a one-layer highly recurrent network of associations, called a Dynamic-Eigen-Net, that we developed to address the limitations of earlier variants of associative nets when scaling up to deal with unstructured input domains like natural language text. We show that spreading activation using a one-hot coded Dynamic-Eigen-Net outperforms the topic model and reaches similar levels of performance as word2vec when predicting human free associations and word similarity ratings. Latent Semantic Analysis vectors reached similar levels of performance when constructed by applying dimensionality reduction to the Shifted Positive Pointwise Mutual Information but showed poorer predictability for free associations when using an entropy-based normalization. An analysis of the rate at which the Dynamic-Eigen-Net reaches asymptotic performance shows that it learns faster than word2vec. We argue in favor of the Dynamic-Eigen-Net as a fast learner, with a single-store, that is not subject to catastrophic interference. We present it as an alternative to instance models when delegating the induction of latent relationships to process assumptions instead of assumptions about representation.

TikTok employs sophisticated algorithms to deliver users increasingly personalised content over time. We investigated the potential for these algorithms to exacerbate eating disorder symptoms by analysing 1.03 million TikTok videos delivered to 42 individuals with eating disorders (76 % anorexia nervosa) and 49 healthy controls over one month. Within this video corpus, we identified four video categories relevant to eating disorder psychopathology: appearance-oriented videos, dieting videos, exercise videos, and toxic eating disorder (akin to "pro-anorexia") videos. Multi-level models predicted the likelihood of users' algorithms delivering these videos and the likelihood of users "liking" (i.e., volitionally engaging with) these videos. Algorithms belonging to users with eating disorders delivered more appearance-oriented (+146 %), dieting (+335 %), exercise (+142 %), and toxic eating disorder videos (+4343 %). Stronger biases in users' algorithms toward these videos were associated with more severe eating disorder symptoms. Whilst users with eating disorders were slightly more likely to "like" these problematic video categories (e.g., dieting videos: +23 % versus controls), their algorithms were far more likely to deliver these videos in the first place (dieting videos: +335 % versus controls). Our results provide preliminary evidence that the TikTok algorithm might exacerbate eating disorder symptoms via content personalisation processes that are desensitised to volitional user actions (i.e., "liking" videos).

Transformer models of language represent a step change in our ability to account for cognitive phenomena. Although the specific architecture that has garnered recent interest is quite young, many of its components have antecedents in the cognitive science literature. In this article, we start by providing an introduction to large language models aimed at a general psychological audience. We then highlight some of the antecedents, including the importance of scale, instance-based memory models, paradigmatic association and systematicity, positional encodings of serial order, and the learning of control processes. This article offers an exploration of the relationship between transformer models and their precursors, showing how they can be understood as a next phase in our understanding of cognitive processes.

Predication is the process by which the meaning of words are altered as a consequence of the contexts in which they appear. Kintsch provides an algorithm to capture this process. The model is based on Latent Semantic Analysis and has the advantage that it relies only on the statistical analysis of word occurrence—and hence provides a coherent and practical account of how representational content might come to be. Similarly, large language models—such as the Bidirectional Encoder Representations from Transformers (BERT) update the initially context-independent representations of words based on their context of use in a data driven way. In a sense, BERT models are a series of predication layers. In this paper, predication, BERT and the relationship between them is explained.

We re-examined whether different time scales such as week, day of week, and hour of day are independently used during memory retrieval as has been previously argued (i.e., independence of scales). To overcome the limitations of previous studies, we used experience sampling technology to obtain test stimuli that have higher ecological validity. We also used pointwise mutual information to directly calculate the degree of dependency between time scales in a formal way. Participants were provided with a smartphone and were asked to wear it around their neck for two weeks, which was equipped with an app that automatically collected time, images, GPS, audio and accelerometry. After a one-week retention interval, participants were presented with an image that was captured during their data collection phase, and were tested on their memory of when the event happened (i.e., week, day of week, and hour). We find that, in contrast to previous arguments, memories of different time scales were not retrieved independently. Moreover, through rendering recurrence plots of the images that the participants collected, we provide evidence the dependency may have originated from the repetitive events that the participants encountered in their daily life.