Longitudinal relational agents rely on memory for continuity, but remembering can lead to intrusion, clinical overconfidence, emotional manipulation, and opacity. We present Relic, a governance architecture for diegetic relational agents, and Gumi, a subject-specific companion agent initialized and governed through Relic. The architecture separates what the system can observe, what researchers can review, what Gumi can remember, and what Gumi can say. We describe three mechanisms. Relic represents sensitive safety signals as researcher-facing governance objects rather than user-facing explanations or agent memories. Shared Continuity Memory organizes relational markers around subject-confirmed wording, corrections, explicit follow-up permissions, and recall limits. We specify a runtime mediation layer for delivery, proactivity, session identity, output transformation, and platform allowlists; adoption across live paths remains partial. These mechanisms aim to preserve relational continuity without turning Gumi into a clinical assistant, mood tracker, general-purpose assistant, or opaque profiling system. We contribute the architecture, object model, researcher workbench design, and a proposed evaluation framework for governed relational memory. An exploratory pilot involved the author as P1 and one external participant as P2. Weekly self-reports from P1 and researcher-led interviews with P2 provide descriptive observations of early use and perceived naturalness. The pilot does not support claims about clinical efficacy, diagnosis, therapeutic outcomes, general usability, or deployment safety. Relic provides research infrastructure intended to make relational memory, safety constraints, and agent behavior inspectable and testable.
How would we know whether a machine truly understands the world? We argue that one key capability is observational theory learning: acquiring reusable theories from pure observations and applying them to new situations. Current visual reasoning benchmarks focus primarily on the latter, presenting each task as a few-shot puzzle grouped by a common rule. We introduce BabyTheorist, a benchmark for controlled training and evaluation of observational theory learning. BabyTheorist generates continual streams of before-after visual observations from hidden compositional programs; during training, learners see only the resulting transitions, with no program labels, task identities, or support sets. To support theory learning and evaluation, BabyTheorist controls the program language, a curriculum over program complexity, recurring frequent subprograms, and a continual train-test protocol with level skipping. Seven representative baselines plateau well below ceiling, with the gap widening as the curriculum advances---evidence that BabyTheorist isolates a capability current program-learning approaches do not yet address.
Synchronous MoE reinforcement-learning post-training suffers from response-length imbalance. Existing methods reduce request tails, yet completed ranks remain bound to Expert Parallelism (EP) collectives and expert ownership. In a representative 16-accelerator Qwen3-30B-A3B rollout, this dummy run traps 32.83% of rank-time. AdaFloor converts this EP-coupled capacity into bounded leases for auxiliary work to improve system resource efficiency through rank-time reuse. Historical prompt-tail prediction and a calibrated guard construct the KV-safe layout. AdaFloor then selects the floor-specific safe quorum schedule with the largest predicted release area. Its runtime shrinks the EP group and leases detached ranks until full-world restoration. Across three updated-policy seeds on DeepScaleR mathematical reasoning rollouts, its capacity-oriented Natural floor2 configuration with TailGuard reduces Vanilla's 29393.9 rank-seconds of dummy run by 96.15%. Rollout shortening eliminates 18693.6 rank-seconds, while coordinated exits convert another 9568.6 rank-seconds into reusable capacity, leaving 1131.7 rank-seconds. For sidecar reuse, its Planned residency floor4 configuration without TailGuard serves 3134 Qwen3-8B GSM8K queries per epoch at 99.46% accuracy and incurs only a 2 to 3% primary throughput reduction against the same configuration with sidecar execution disabled. Its latency-oriented Planned residency floor4 configuration with TailGuard reduces rollout time by 24.09% relative to Vanilla and by 2.31% relative to LengthSort with TailGuard. On public Eurus-2-RL code rollouts with the same model and 16-rank topology, AdaFloor reduces rollout time by 20.56% and dummy run by 89.13% while generated work and mean reward remain stable.
Oscillatory phase coding (OPC) is widely observed in neural systems and has been implicated in cognitive functions such as memory and sequence processing. However, its computational role and neural substrates remain largely unclear. Here, we studied OPC by training recurrent neural networks (RNNs) with adaptation to perform sequence working memory tasks with variable delays. After training, we found that OPC emerges in the RNN spontaneously: the memorized items recur at distinct phases of a shared rhythm, with oscillation amplitude encoding item identity and phase encoding serial order. Ablation studies identified three key elements required for OPC: adaptation in neural dynamics, variable delay in memory retention, and flexibility in reading out information. Neuron-type analysis revealed the emergence of content and time cells in the trained RNN, and they process item identity and sequence order separately. Recurrent connectivity analysis revealed phase-organized connections among time cells and structured connections within digit-selective content-cell groups. Inspired by these findings, we proposed a neural circuit model which reproduces OPC through interactions between content units and an adaptive time ring. Together, this study reveals how OPC arises from network structures and task demands, offers insights into the implementation and computational mechanisms of OPC in biological systems, and inspires us to develop computational models for human-like cognitive functions.
At work, people develop skills through experience: novices acquire skills by completing tasks, while experts retain theirs through continued practice. AI assistance changes this process by offering workers a choice between solving tasks independently and accepting the AI's answers. While relying on AI saves immediate effort, it can also displace the practice needed to acquire and retain skills, including those required to oversee the AI. We extend an existing accept-or-solve model to a sequence of tasks in which workers' skills evolve through independent practice and reliance on AI. We compare myopic workers, who maximize immediate utility, with strategic workers, who also value their future skills. We show that novices can become trapped at low skill levels, while experts may lose skill, even as AI-assisted outcome accuracy remains high. Strategic workers, by contrast, may choose to solve independently even when AI is more accurate, recognizing the future value of acquiring and retaining their skills. We then study incentives that reward correct independent work and characterize the payments needed to encourage skill acquisition and retention. We show that independent work can become self-sustaining once workers acquire sufficient skill, eliminating the need for continued incentives. Counterintuitively, supporting a higher skill target can sometimes reduce long-run incentive costs, either by enabling self-sustaining independent work or by lowering the payments needed to maintain proficiency. Our findings thereby shift the focus from whether workers should rely on AI on individual tasks to how organizations can sustain human expertise over time, and which skills are worth investing in.
A central challenge in AI is specifying objectives that align with human preferences. While existing theoretical results elucidate how misspecification impacts performance, they typically do not distinguish between merely suboptimal and catastrophic outcomes. To enable rigorous analysis of catastrophic risk from AI misalignment, we introduce a formal definition of catastrophic performance. Our definition is based on two reference values. The first, baseline value, captures the value attained if no agent is deployed. The second, orthogonal value, captures the value of a trajectory that is unrelated to human preferences, in a sense we describe. Orthogonal value is typically far worse than baseline value. We interpret these reference values through examples in which past human activity has altered the state of the environment to human benefit. Starting at the resulting state, an AI agent is deployed with the goal of accelerating further progress. Catastrophic performance represents this agent's undoing of benefits afforded by human activity.
LLM-based coding agents are primarily evaluated by their solve rates. However, this metric does not reveal how difficult the solved tasks are or account for differences in task difficulty across benchmarks. Determining the difficulty of individual tasks by repeatedly running multiple agents is costly, while a predicted difficulty score alone provides limited guidance on what support an agent needs. To address these challenges, we propose DARE (Difficulty Analysis for Repair Enhancement), which integrates repair-task difficulty measurement, prediction, feature-group contribution analysis, and task-specific repair support. We define continuous task difficulty as the observed failure rate across multiple coding agents. The DARE difficulty model predicts this difficulty using 151 static features extracted from benchmark artifacts. The framework also analyzes the contributions of seven Repair-Task Feature Groups to identify the characteristics most salient to each task and uses this information to select task-specific agent support. We evaluate DARE on four benchmark partitions comprising 2,594 tasks. The difficulty model's aligned predictions achieve Spearman's $\rho$ = .476-.621 on held-out partitions, and the most salient Feature Groups vary across tasks and benchmarks. Furthermore, coding-agent experiments show that matching support to each task's salient Feature Group raises repair success from 54.7% to 79.6% on nine high-difficulty tasks. These findings demonstrate that difficulty prediction and interpretation can lead to practical improvements in automated program repair (APR), while also providing a basis for difficulty-stratified agent evaluation, cross-benchmark comparisons, and identifying areas of repair capability that warrant further improvement.
Scientific agents can find interventions that work without understanding why they do. We study this failure mode in controlled synthetic regulatory systems, where agents must experimentally identify a hidden fault and rescue the resulting phenotype. Across seven LLMs, every model restores the healthy phenotypes more often than it identifies the fault. Claude Opus 4.6 and Sonnet 4.6 rescue every organism while diagnosing only 55% and 35% of them. This failure is not simply due to insufficient exploration or unavailable evidence. In most cases, the cause is experimentally identifiable, and agents encounter evidence that distinguishes competing causes in 91% of relevant sessions, and still frequently misdiagnose the system. Targeted probes reveal where experimental reasoning breaks down. Models reliably interpret decisive evidence and revise hypotheses when the causal comparison is explicit, and can locate that evidence within a full experimental record. However, they struggle to predict the outcome of the experiment under each competing cause, a necessary step in using the observed result to distinguish between them. Agents can therefore gather the right evidence and still infer the wrong cause. As AI systems take on a growing role in synthesising scientific evidence, these findings highlight the need to evaluate how they use evidence to reach causal conclusions, alongside whether their interventions succeed.
Video games provide a practical setting for collecting vision-language-action (VLA) data, whose acquisition in physical environments is costly. Existing datasets, however, remain limited by restricted game coverage, sparse language supervision, or temporal hierarchies that do not necessarily reflect changes in task objectives. We introduce AceAction, a dataset of human gameplay across multiple 3D games for learning skill composition and subgoal transitions. AceAction organizes demonstrations around complete tasks, linking overall instructions, ordered subgoals, and local skill instructions to temporally aligned visual observations and keyboard-and-mouse actions. To construct these annotations, we develop a vision-language model pipeline that identifies subgoals from changes in objectives or interaction targets, anchors navigation goals to visually recognizable waypoints, and composes adjacent subgoals into task instructions. The resulting demonstrations cover navigation, reconnaissance, combat, interaction, and recovery, with explicit stage boundaries that connect task progression to fine-grained control. We assess annotation quality through model-based verification and human inspection. An audit of 500 trajectories yields a 95.8% pass rate across independent reviewer assessments. AceAction provides a data foundation for studying how agents execute, combine, and switch skills to follow complex instructions in dynamic game environments.
An experimental agent can succeed for the wrong reasons. Evaluations, however, often focus on the outcome, not the experiments, evidence, and inferences that produced it. We introduce LUCA, a configurable synthetic world for evaluating this process under fully known ground truth. Agents learn unknown regulatory systems through budgeted experimentation, then predict unseen interventions, diagnose hidden faults, and attempt rescues. This setup allows us to evaluate experiment selection, prediction, diagnosis, and intervention separately. Across seven frontier LLMs and three baselines, these capabilities dissociate sharply. Agents frequently rescue malfunctioning systems while diagnosing the wrong fault; diagnosis varies 7x across models while prediction varies only 1.4x. Models also adopt surprisingly different experimental strategies, and greater exploration does not reliably yield better performance. In one failure mode, agents often perform the experiment needed to distinguish competing causes, but still fail to use the resulting evidence. LUCA turns these hidden failures into measurable behaviours, providing a controlled environment to study where experimental agents succeed and fail. Code: https://github.com/RelationRx/luca.
Speculative decoding accelerates vision-language model (VLM) inference without altering the output distribution, yet two factors limit its gains. Existing drafters either lack the image evidence required for visually grounded tokens or pay to condition on every visual token, although a short draft block typically depends on only a few image regions. Moreover, the value of verifying additional candidates varies with GPU load, so no fixed verification structure is efficient across serving conditions. We present Uni-LaSpec, a block-parallel speculative decoding framework that addresses both factors. To improve draft quality, Uni-LaSpec reuses the verifier's own attention: the query retained from the preceding verification ranks the cached, int8-quantized visual keys, and a cross-layer vote selects a few visual tokens that the drafter fuses through gated cross-attention. Acceptance-aware training further weights each draft position by its contribution to the accepted prefix. To allocate verification, a marginal-cost criterion selects the structure per request and per round: draft trees exploit idle verifier capacity under light load, whereas a dynamic head truncates linear drafts under heavy load. Across six VLMs from three model families and 17 text and multimodal benchmarks, selecting eight visual tokens retains at least 98.6% of the acceptance obtained from the full image at sub-millisecond selection cost. With greedy decoding at concurrency one, Uni-LaSpec achieves up to 5.11x text and 3.24x multimodal throughput over autoregressive decoding. Across six load levels and three GPU types, it outperforms autoregressive decoding in every setting and matches or exceeds the best fixed verification strategy in 123 of 144 settings; under bursty arrivals, it reduces mean end-to-end latency by 47.6% relative to linear verification.
Diffusion models and more recently Flow models have widespread application. In this work we consider degenerate Diffusion models in which the per-step generator is deterministic. We construct a modified variational framework and show a formal equivalence between Flow models and Degenerate Diffusion models. For our discrete-time variational objective we show that the optimal degenerate generator recovers the data distribution in the limit of an infinite number of generator steps. Our work therefore provides a unified perspective, showing that discrete-time implementations of flow matching and the discrete-time Degenerate Diffusion model can be considered within the same simple variational framework, removing the need for the continuous-time machinery normally used to justify flow matching models.
Data flywheels promise continuous improvement via insights from deployment data. Understanding how the deployed model succeeds and fails can steer subsequent data collection towards patching failures and reinforcing successes. However, robot deployments create large volumes of high-dimensional rollout data, which is challenging to understand and confidently ground data collection decisions in. To make deployment data easier to analyze and act on, we propose behavior graphs, which compress policy rollouts into a single interpretable, directed graph. We construct the graphs by learning to identify similar behavior modes across rollouts, connecting the behaviors temporally, and finally linking them to outcomes. In a user study on three real-world tasks, where participants analyzed rollouts to choose a data collection strategy, 16 of 18 preferred behavior graphs to rollout videos, reaching comparable conclusions with 40.3% lower perceived temporal demand and 37.6% lower effort. Beyond interpretability for deployment data, behavior graphs are capable progress predictors: across 43 datasets, they improve the rank correlation between predicted progress and true temporal order by 22.9% and reduce mean absolute error by 53.5% relative to the strongest baseline. As the number of deployed robots grows, behavior graphs lend structure to high-dimensional policy rollouts and make deployment data easier to interpret and reason about.
We introduce Radial Kernel Attention (RKA) for long-range modeling in polymer machine learning force fields. Existing approaches often rely on local neighbor graphs, which limit communication across polymer chains and their surrounding environments, whereas explicit all-pair attention incurs quadratic computational cost. RKA maps interatomic distances from the cluster centroid to learned radial features and combines their kernelized overlap with content-dependent queries and keys. Factorized aggregation and readout enable global feature exchange at linear complexity in the number of atoms for fixed feature dimensions. To ensure physically consistent predictions, we insert the invariant scalar update between EquiformerV3 local blocks and obtain conservative forces by differentiating the total energy, including its dependence on radial coordinate . Our models are trained on official OPoly26 training set for 12 epochs and evaluated directly after training on the complete official validation set. A matched 5M comparison shows that RKA reduces per-atom energy and force MAEs by 67.4% and 36.3%, respectively. The 20M RKA model attains a total-energy MAE of 15.81 meV (per-atom MAE of 0.1182 meV/atom) and a force MAE of 2.619 meV/Angstrom,improving all three metrics over UMA-S-1.2 and UMA-S-1.2.1 on the same validation structures. RKA provides a compact pathway for information exchange beyond local graph connectivity in polymer machine learning force fields.
Vision-language-action models (VLAs) map language instructions and visual observations to robot actions, showing strong capabilities across robotic manipulation tasks. However, in our probing experiment, we find that action representations near the action output do not fully preserve task-relevant visual grounding. Motivated by this late-layer grounding gap, we introduce STAR, for STeering Action Representations, a training-free inference method for pre-trained VLAs. STAR applies an exponential moving average over late-layer residual updates, allowing later action representations to inherit object-aware update directions from earlier layers without model updates or architectural changes. It further uses recent proprioceptive states to adapt the mixing strength, using conservative mixing when execution dynamics change rapidly. Experiments on robot manipulation benchmarks show that STAR improves late-layer object awareness and increases average task success rates by 4.0 percentage points. Our code is available at https://github.com/PKU-VaLuE-Lab/STAR.
Recommendation-driven platforms can reshape not only what users see, but also who becomes visible and how community boundaries are negotiated. We examine Xiaohongshu discussions surrounding a perceived World Cup-related audience influx to understand how users attributed community change and how those explanations related to boundary-management strategies. We analyzed 262 comments from eight high-engagement threads using researcher-led thematic analysis, structured coding, and permutation-based association analysis. Users described personalization as both social separation and a boundary that could fail, interpreted platform growth through audience acquisition and commercialization, and contested whether gender should define community membership. Among 70 comments with explicit attributions and strategies, attribution and strategy were strongly associated (Cramer's V = .583): algorithmic attributions concentrated in technical boundary management, while social-actor attributions more often aligned with social enforcement or normative socialization. These findings connect causal sensemaking with boundary management and show how audience restructuring can create governance consequences on recommendation-driven platforms.
Visual counting benchmarks report whether vision-language models produce the correct number, but not whether errors arise from perception or reasoning. We introduce DecoyCount, a benchmark built from 1,586 real-world images with designed decoys across seven mechanisms, along with paired prompts: a neutral question, a matched alignment cue, and a novel multi-hop, multi-image question that situates the same counting problem in a realistic scenario. Using this design, we find that counting failures are induced partly by visual misalignment: models can execute the intended counting process when cue-guided but fail to select it unaided, and we term this difference in accuracy the reasoning-alignment gap. Despite being highly capable on general reasoning, frontier models remain below 50% accuracy on our benchmark, while alignment gaps range from 12.1-22.1 percentage points; strikingly, we observe that more capable models, greater thinking budgets, and agentic processes often induce larger gaps. Because humans may bring context that models lack, we test whether textual underspecification contributes to misalignment using a scenario-based multi-hop condition without explicit hinting, but this added context does not improve aggregate performance. Our analysis concludes that stronger general reasoning does not by itself generalize to aligned visual reasoning, leaving misalignment with human visual expectations as a substantial component of the remaining deficit. DecoyCount and the evaluation code are released at https://github.com/elorian-ai/decoycount.
Life experiences can shape policy attitudes, but we only live one life. Citizens may therefore fail to support policies addressing risks they have never personally experienced. Narratives can make such risks immediate and personal, but personalized stories remain difficult to produce at scale. Here we ask whether AI can make distant scenarios feel concrete by simulating a participant's life as if that scenario occurred. We present the As-If Machine (AIM), a multi-agent system through which people co-write personalized counterfactual narratives---stories about their own lives under different policy futures. In two pre-registered and quota-matched experiments on science funding (N=1,007) and satellite cybersecurity (N=894), we compared AIM with two controls: reading an expert-written vignette on the same topic and co-writing with AIM about an unrelated topic. AIM increased psychological closeness---how close a risk feels to the self---by nearly identical margins in the two domains (0.36 and 0.38 standard deviations (SDs) against the vignette, 0.80 and 0.90 against unrelated writing). AIM increased psychological closeness even among participants with unfavorable views of AI. For policy attitudes, AIM's effects were 0.16 SDs against the vignette and 0.10 SDs (n.s.) against unrelated writing for science funding; the corresponding effects for satellite cybersecurity were 0.54 and 0.67 SDs. Across 23,496 human-AI autocomplete interactions, participants and the system played distinct roles: AIM supplied concrete detail that grounded the scenario, and participants supplied first-person perspective and emotion. Co-writing with the As-If Machine makes distant policy risks feel close.
HarnessMinusMinus (HarnessMM) is a research project on improving the efficiency of agent harnesses, the software coordinating model reasoning, tool execution, and feedback. This technical report consolidates its shared framework, implementations, and empirical studies across coding, tool-based, and visual-agent settings. Agent harnesses often invoke reasoning after each tool result, even when earlier calls have already determined subsequent actions. HarnessMM addresses this overhead through guarded decision reuse and recursive self-improvement (RSI). RSI refines author prompts, context settings, and adapter rules using training feedback, with validation selecting deployment configurations. At inference time, a controller reuses resolved decisions while keeping actor weights fixed. Executable Decision Graphs encode actions, arguments, dependencies, and guards. Guarded Decision Reuse activates eligible decisions from observed feedback, while Certified Callback Substitution executes them through native tool interfaces, replacing redundant reasoning callbacks. The original actor resumes when new reasoning is required or reuse conditions are unmet. On HumanEval, HarnessMM preserves Codex success at 90.82\% while reducing tokens per task from 66.72k to 47.02k (29.54\% saving). For DeepSeek Harness, success improves from 71.43\% to 85.71\% while tokens decrease from 96.28k to 11.62k (87.93\% saving). An MBPP ablation shows 26.91\% lower cost than single-use graphs and 19.52\% lower cost than retained callbacks. The report also summarizes the HarnessMM-V visual-agent study and documents optimization overhead and execution time where measured. Together, these studies investigate guarded decision reuse across diverse agent harness settings. Code is available at https://github.com/YuhuaJiang2002/HarnessMM.
RNA molecules often function through multiple experimentally observed conformational states. A central challenge is to capture ensemble diversity without sacrificing structural fidelity: samples may collapse toward averaged representative structures or drift into unsupported conformational regions. We introduce REFLEX (RNA Ensemble generation calibrated by FLEXibility), a sequence-conditioned generator for RNA backbone-frame ensembles based on a heteroscedastic stochastic bridge on SE(3). During training, experimental B-factors calibrate residue-wise bridge widths, exposing each observed conformer through local perturbation neighborhoods with residue-specific scales while keeping constrained regions sharper. A learned flexibility condition guides the bridge-field network to recover target conformer-specific frames from these calibrated noisy states. At inference, REFLEX uses only the input sequence, requiring no MSA-subsampling heuristics. On a multi-conformer RNA benchmark, REFLEX achieves the best precision-recall trade-off for ensemble coverage across evaluation thresholds, while recovery-stress ablations indicate gains in target-frame recovery under stronger perturbations.