Humans and AI systems increasingly act within the same social and economic environments. Alignment, the goal of ensuring that AI systems act in accordance with human values, is widely regarded as a central challenge for ensuring that AI benefits society while avoiding harm. We argue that this endeavor is misguided. First, alignment is not well defined, as it presumes a sufficiently stable and identifiable set of human intentions or values. Second, where a concrete notion of alignment can be specified, it is not solvable. Third, where alignment appears solvable, solving it would not necessarily mitigate the negative consequences of AI. A key reason is that alignment is typically framed as a property of individual AI systems, while the resulting hybrid societies of humans and machines are complex systems shaped by interactions and feedback loops. Emergent risks, and not only those posed by individual misaligned systems, are therefore likely to become a major source of harm. We conclude that AI safety is an ongoing challenge and not a problem that can be solved once and for all by verifying a particular model property.
We study risk control for online policy learning, where an agent adaptively collects data and aims to learn a new policy whose risk stays below a tolerance threshold at every round with high probability. Such policy updates and data collection induce feedback covariate shift (FCS): observed data depend on history through evolving policies and the data-collection process. FCS brings two substantive problems for risk control: 1) efficient risk evaluation for each candidate policy and 2) valid policy selection with high-probability risk control. To address these, we propose a permutation-based weighted risk evaluator with weights efficiently approximated by the Sinkhorn algorithm and select policies whose estimated risk upper confidence bounds (UCB) fall below the target threshold, with the UCB obtained through a proposed Gaussian bootstrap method. Theoretically, we prove that our risk evaluator has a no-larger variance upper bound than that of the standard importance-weighted evaluator. Our policy control method also yields an asymptotically valid simultaneous risk bound and provides asymptotic high-probability risk control for the selected policy. Empirically, we evaluate our method on the online contextual bandit problem, safe question answering benchmark, and Medical Claim Filtering tasks. Our method maintains the target risk level, achieves lower risk than baseline methods, has lower variance in risk estimation, and achieves reliable policy selection with high-probability risk control. Further ablation studies support our theoretical claims.
Effective search and rescue (SAR) operations require autonomous systems capable of coordinating multiple heterogeneous agents, adapting to dynamic hazards, and aligning with human mission intent. Existing language-guided reinforcement learning (RL) methods primarily focus on single-agent instruction following and lack mechanisms for team-level coordination in complex, partially observable environments. This paper presents a Cognitive Multi-Agent Framework based on Language-Grounded Hierarchical Reinforcement Learning (LG-HRL) that integrates natural-language understanding, hierarchical policy decomposition, and a shared cognitive workspace to achieve cohesive multi-agent behavior in SAR scenarios. Human instructions are interpreted by a large language model (LLM) and transformed into structured semantic cues that guide high-level strategy selection and role assignment across UAV and UGV agents. A formal Language-Grounded Multi-Agent Hierarchical Decision Process (LG-MAHDP) is introduced to model the interplay between semantic grounding, hierarchical control, and distributed coordination. Comprehensive experiments in a simulated disaster environment show that LG-HRL improves learning speed by 2.3 times, increases coordination coherence by 37%, and enhances subgoal accuracy and hazard avoidance by 46% and 31%, respectively. The system achieves 20-30% higher victim rescue rates, demonstrates emergent role specialization, and adapts immediately to human verbal updates. Ablation studies confirm that both the shared cognitive workspace and language grounding are critical to overall system performance. Overall, this work establishes a new paradigm for language-guided multi-agent SAR robotics, demonstrating that semantic grounding combined with hierarchical reasoning enables safer, more efficient, and more interpretable autonomous rescue operations.
Vision-language-action (VLA) models now exceed 95% success on the LIBERO simulation benchmark, yet it remains unclear how much of this performance requires generalization beyond the training demonstrations. In this work, we show that retrieving nearby demonstrations and reusing their action chunks achieves a success rate of 91.7 % on LIBERO without policy training. Under LIBERO-Plus perturbations, retrieval shows the smallest drop of all evaluated policies. Across nine publicly available VLAs, fine-tuning mostly makes the vision-language backbone a better retrieval key, and the actions of the VLAs remain close to the demonstrations. In addition, we introduce Local Action Consistency (LAC), an offline metric that measures whether nearby observations in representation space are also associated with similar future actions. We show that LAC tracks the success rate of retrieval and can guide the training of a retrieval key. Our evaluation and source code are available at https://github.com/sereact/lerobot_knn.
Multimodal large language models (MLLMs) are typically evaluated by whether they produce correct answers, but correctness alone does not reveal whether their outputs are supported by the right visual evidence. We introduce VIGOR, a benchmark for evaluating VIsually GrOunded Rationales in MLLMs. VIGOR represents visual rationales as explicit concepts paired with pixel-level groundings, covering multiple abstraction levels including low-level visual properties, intermediate structures, objects, parts and high-level motion, relations. To construct VIGOR at scale, we develop an automatic annotation framework that reuses existing dense supervision when available and generates missing concept groundings through proposal-based annotation followed by conservative multimodal verification. Using VIGOR, we evaluate whether MLLMs "look at what they say" by comparing model-derived visual evidence for each expressed concept against annotated ground-truth rationales. Our results show that current MLLMs exhibit limited visual rationale correctness and reveal distinct grounding failures across concept levels. VIGOR provides a benchmark and scalable data construction pipeline for studying whether MLLM outputs are supported by appropriate visual evidence. Code is available at https://github.com/deep-real/VIGOR
Attention, the core of language models, is a digital calculation over tokens: an exponential selects them, a division normalizes and a positional code sets the order. We show that attention can instead emerge from the synchronization of a receiver with its context tokens. Each past token is an oscillator whose state carries its content and whose frequency drifts with its age. The present token selects them by content, aligning an antenna with its query, and by position, setting the frequency of a receiver with its dial. Solving Kuramoto's problem for this receiver yields an attention layer in closed form with two phases: above a critical coupling set by the context, the receiver locks and its weights normalize; below it, it stays silent and its weights sum to less than one. Trained networks use both phases to select, abstain and count, fail where the theory predicts and keep reading far beyond the training length. Each part of attention finds its place in one device, a radio head: it tunes in to what to hear and when, and stays silent when nothing relevant is on air.
Effective teaching requires reasoning about the learner as well as the task, a capacity that Theory of Mind describes as inferring another agent's latent state from its behavior. We ask whether a stronger LLM teaches a fixed student better when it applies this principle. The student is an LLM agent acting through a harness that fixes its tools, context, and action budget, so a procedure that helps one model can exhaust the budget of another. We introduce Student-Aware Teaching (SAT). The teacher builds a learner model that pairs an empirical profile with hypotheses about the student's limitations, tests these hypotheses with diagnostic probes, and writes reusable skills for the diagnosed failures. A skill enters the library only after it raises success on matched development tasks. With Claude-Sonnet-5 teaching DeepSeek-V4-Flash, SAT reaches 86.46% task success across four primary runs, leading both a profile-free teacher with the same framework (81.04%) and the untaught student (75.00%) in each run. Removing the analytical learner model lowers both prediction and success, and under misleading learner profiles none of the 48 candidate skills passes admission. When DeepSeek-V4-Flash and Gemini-3.8-Flash exchange libraries, each performs better with its own. Saved libraries keep average gains of up to 11.1 points after the teacher leaves. The benefit depends on a student that can follow instructions and a deployment route that delivers them.
Coding agents built on large language models, run in automated improvement loops, can discover more efficient algorithms. We present AIOA, a framework that brings this approach to the numerical algorithms of atomistic simulation: an agent improves an algorithm against a pre-specified metric in a loop guarded by verifiers it cannot alter. Applied to ASE, a widely used atomistic simulation library, AIOA targets two facets of an algorithm: its implementation, rewritten to run faster with unchanged results, and the method itself, redesigned to converge with fewer force evaluations. Agents made six kernels 22-24 times faster on average in Python with NumPy and SciPy. On unseen problems, optimizers from the best configuration needed 3.1-3.4 times fewer force evaluations than L-BFGS for structure relaxation and 4.6-17 times fewer than standard climbing-image nudged elastic band calculations for reaction paths. The agents' optimizers also contain two algorithmic components we did not find in the literature.
Learning-based decentralized collision avoidance must remain reliable under state-estimation uncertainty and heterogeneous agent dynamics. Existing methods typically address these factors through training-domain variation or by exposing uncertainty and capability estimates to the policy at runtime. However, both leave the policy to infer how these quantities should alter the physical separation required for safe interaction, which can couple collision-avoidance behavior to relationships learned from the training distribution. We introduce the Dynamic Oriented Bounding Box (DOBB), an online geometric representation that maps the current state estimate, uncertainty, and agent-specific response characteristics into a dynamic, direction-dependent safety margin. We derive geometric features that jointly characterize the relative configuration and severity of interactions between agents' DOBBs, along with a reward function that preserves the physical interpretation of those interactions. Across four evaluation environments, DOBB achieves the lowest mean total failure rate of 2.6%. This performance persists under out-of-distribution shifts in actuation dynamics and state-estimation uncertainty, with collision-based failure remaining at or below 2.3% in every tested condition. We attribute this robustness to the geometric abstraction of the safety margin, which is less coupled to the specific uncertainty and actuation regimes encountered during training.
Failure taxonomies for LLM reasoning are increasingly reported as measurements: papers state what fraction of a model's errors fall into named categories, then compare those fractions across models. We show that for one such category the reported fraction is largely a property of the measurement procedure. We defined Compositional Collapse (a model notices a required sub-part of a problem, then drops or misapplies it) and measured it across twelve models and eleven languages on a public multilingual math benchmark (2,316 graded records). First, 50.1% of failures carry evidence for more than one category, so the reported plurality depends on the order in which the classifier tests them: reversing two tests moves it in eleven of twelve models, and across all six orderings Collapse's share spans 26.4% to 76.4%. Second, agreement collapses for the category but not for the task. Native speakers of all eleven languages judged 179 transcripts, agreeing with our inspector at kappa = 0.932 on answer correctness but only kappa = 0.093 on Collapse. A second native speaker per language re-rated 142 items; on 121 determinate pairs the two humans agree with each other at kappa = 0.659, while each agrees with the inspector at kappa <= 0.094. Re-running the same inspector model with the human instructions verbatim raises kappa against Human 2 from 0.147 to 0.636, so the gap is largely attributable to the prompt, not the model. Third, the inspector labels Collapse 1.81 times as often as native speakers, but only 1.17 times once a defective benchmark item (a deleted clause in the Bengali and Chinese translations) is excluded. We recommend four cheap checks before comparing failure-mode rates across models: report label co-occurrence, permute the tie-break, validate on a shared independent sample, and audit systematic disagreements.
Optimization methods, including the Muon optimizer based on the Linear Minimization Oracle (LMO) and momentum orthogonalization, have shown strong empirical performance in large-scale and distributed learning, yet their robustness to adversarial behavior and data heterogeneity is not well understood. We study these methods within a unified first-order trust-region framework and analyze Muon-based updates with robust aggregation for Byzantine-robust distributed and federated learning. Our analysis does not rely on the i.i.d. assumption and introduces a non-Euclidean notion of heterogeneity consistent with the trust-region geometry. Under this framework, we establish convergence guarantees for both general non-convex and star-convex objectives. Our results provide a theoretical characterization of Byzantine-robust Muon-based optimization, with an additive heterogeneity error floor consistent with known impossibility results, under explicit aggregation assumptions. Under an explicit additive oracle-error condition, the guarantees also cover inexact orthogonalization, with the non-convex Byzantine terms in the bound's numerator unchanged but a smaller progress denominator. In simulated federated training with ResNet-18 and ViT-Tiny on CIFAR-10, additional image benchmarks, and a 38M-parameter byte-level language model on TinyStories, the tested Muon configuration attains higher attacked accuracy (lower perplexity) than the tested SGD and normalized-SGD configurations, and than sign updates on CIFAR-10. Fixed-honest controls further show a substantial RFA advantage over averaging under a bounded RFA-query attack, higher attacked accuracy for Muon-NS than tensorwise Euclidean normalization from shared checkpoints, and lower held-out TinyStories perplexity than a tuned AdamW baseline.
On-policy distillation (OPD) reduces train-inference distribution mismatch by supervising a student on its own trajectories. For multi-turn tool-use agents, conditioning the teacher on a validated same-task successful trajectory as a reference can provide useful hindsight, but different successes may reach incompatible interaction states and offer misleading supervision. We cast successful-reference selection as a free-target optimal transport problem, which reduces to assigning each failed trajectory its most compatible successful reference. Because latent state paths are hidden, we use action-sequence alignment as an observable proxy for compatibility and instantiate it with longest common subsequence (LCS), which identifies a shared ordered backbone of actions while allowing recoverable detours. This perspective makes trajectory compatibility the central criterion for reference selection in multi-turn OPD. Experiments on ALFWorld and the airline domain of $\tau^3$-bench show that LCS is the only evaluated reference selector to improve over reference-free OPD on both ALFWorld overall scores and all three airline metrics. Without further optimization, the airline-trained models also transfer to the retail domain, where LCS improves all three metrics.
Retrieval-augmented language model systems generally select context by query-conditioned relevance. We argue that long-lived AI systems require a second operator: activation-conditioned memory availability. We specify an architecture in which request and goal cues seed an authorized memory graph, activation and fan attenuation allocate a bounded traversal frontier, retrieved entities recursively emit cues, and a typed evidence graph is reconstructed before generation. The specification yields five architectural consequences about candidate selection, traversal control, graph regularization, evidence reconstruction, and reinforcement. We derive implementation-level cost estimates and define ablations, failure conditions, and a reproducible evaluation protocol. The central claim is narrowed or falsified if a simpler graph reranker matches full traversal on causal and temporal recovery under equal budgets. This architecture-and-analysis preprint reports no benchmark measurements. If the claim holds, retrieval returns the evidence path that makes a memory available, not only a ranked list.
Recent work on latent-free diffusion transformers has demonstrated the feasibility of training image generators without a latent VAE. This avoids compression artifacts from the VAE and simplifies the training pipeline. In particular, JiT showed that even a standard transformer can be trained to generate images when the model is trained to predict the final image ($x$-prediction). Even though JiT unlocks training efficiency comparable to latent models, it still suffers from reduced image quality likely arising from modeling the image as a collection of large patches. In this work, we revisit JiT and show that adding simple convolutional pixel-pathway layers in a post-training phase significantly improves generation quality. Moreover, we show that the added pixel-pathway does not need to be active in the early denoising stages, and can be activated as late as in the last $20\%$ of the denoising process. Our experiments on ImageNet@256 show improvements compared to the original JiT results, achieving an FID of $1.78$ (from $2.34$ with corrected sampling), without using additional losses.
System identification requires expert decisions on the identification method, incorporation of prior physical knowledge, and model validation. We propose ASPIRE, a framework for nonlinear system identification in which a large language model agent selects and executes the identification method. ASPIRE constructs closed-form, nonlinear static or dynamic models from measured data, optionally supplemented by a system description, the intended use of the model, and physical requirements expressed formally or in natural language. Each model includes its parameters, an estimation record, an interpretation of its terms, and an error evaluated on test data inaccessible to the agent. Formal physical requirements are verified independently. ASPIRE operates autonomously or interactively, allowing the agent to ask questions and the user to request further analyses and identification rounds. We evaluate ASPIRE on three measured systems: a mobile robot, a rotational system with dry friction, and a laboratory fermenter. Each system is tested with and without a description, in five runs per task. All 30 runs delivered valid models, at a cost of 0.96 to 3.38 US dollars per run. An assessment by the authors found that, for the 30 models recommended by the agent, one per run, the interpretation of the model and the documentation of its structure selection and parameter estimation fully satisfied 132 of 150 assessed items and partly satisfied 18. The most frequent discrepancy, found in nine models, concerned inaccurate or unsupported claims about model structure selection.
In the absence of extrinsic rewards or task supervision, human behavior is guided by intrinsic motivation---neural signals that drive curiosity, creativity, play, and other essential facets of intelligence. Intrinsic motivation plays an important developmental role in building robust world models that enable generalization and adaptation. Yet existing studies leave open how a broad range of intrinsic motivations interacts with world-model representation learning to support downstream control. This limits our understanding of which intrinsically collected experiences build reusable models. To address these gaps, we integrate 24 intrinsic objectives, including established methods and explicit variants spanning curiosity, entropy, and competency, into a common model-based reinforcement learning framework. We evaluate the world models they build through zero-shot, few-shot, and multi-task control in locomotion, manipulation, and open-ended gaming environments, using three representation-learning strategies: observation reconstruction, latent prediction, and redundancy reduction. Curiosity-driven exploration with observation reconstruction yields the strongest aggregate family profile, although the best individual pairing depends on the experiment. In compositional vision-based tasks, intrinsic motivation derived from both episodic and lifelong novelty---providing a complementary memory to the world model's generic recurrent state---achieves the best zero-shot transfer and few-shot adaptation. These findings motivate designing intrinsic motivation and representation learning jointly. Our results demonstrate world models as a powerful tool for measuring the quality of intrinsically motivated experience, and positions intrinsic motivation as a key ingredient for building world models autonomously.
Large language models (LLMs) are increasingly trusted with mathematical, scientific and engineering work, yet their mistakes are often confidently written and hard to spot. A single wrong step tends to propagate, making the model waste inference tokens following a false premise. Existing safeguards either check only the final answer and restart from scratch, or rely on a verifier LLM that competes for GPU capacity and slows down response time. We propose online, deterministic verification: many kinds of reasoning steps can be validated by code, so a step can be verified in real time, as it is written, without any model call or GPU use. Whenever a check fails, we point out the mistake and let the model continue from the same spot. We demonstrate this approach on symbolic mathematics, where derivations already contain explicit expressions that a computer algebra system can check in milliseconds. Tested on the ASyMoB benchmark with three open-weights models, our online validation improves accuracy by 1.4-5.6 percentage points, corresponding to a 1.8-14.9% net reduction in incorrect answers, while increasing token usage by only 6-7%. Because the verifier is independent of the model, the approach complements progress in LLMs rather than competing with it: every new model can benefit immediately, and the benefit grows as deterministic checkers extend to more kinds of reasoning steps.
Inference-time sampling is a central mechanism for improving large language models, with temperature controlling the tradeoff between reliability and diversity. While recent work suggests that varying temperature across samples can improve multi-sample inference, simple fixed-temperature schemes remain the standard in practice. We investigate this discrepancy in the canonical pass@K setting, asking when temperature diversity is useful, and when a single tuned temperature suffices. To this end, we analyze per-question success probabilities across multiple models and datasets, tracking how they vary with sampling temperature. We identify several consistent properties of these profiles, including that harder questions are less likely to share efficient sampling temperatures, although this trend is only logarithmic in hardness. We then derive an efficient method for evaluating the theoretically optimal multi-temperature allocation scheme, and show that although such a scheme improves performance, its gains over single-temperature baselines are typically modest when temperature is tuned as a function of the sampling budget K. We explain this phenomenon through mathematical mechanisms that we call dulling effects, and show how they ensure the robustness of temperature choices across different question distributions. Our results clarify when multi-temperature sampling is useful, and how temperatures should ideally be chosen.
Brain networks exhibit modular organization, where community structures vary substantially across individuals and diseases. Understanding these individual differences is crucial for advancing both basic neuroscience research and precision medicine. However, widely used predefined group-level brain atlases often overlook this heterogeneity. While individualized functional parcellations capture subject-specific variability in community boundaries, they often require lengthy scan times to achieve fine-grained resolution and complicate cross-subject comparisons. To address this trade-off and investigate how individualized community organization captures disease-associated changes across individuals, we propose IC-GNN, an atlas-anchored model that forms distinct, individualized functional communities directly over atlas-defined brain regions. IC-GNN derives discrete, individualized multiscale communities from resting-state functional magnetic resonance imaging (fMRI) data, refines community-guided connections, and adapts shared regional weights to each subject's community context to enable both individualized modeling and group-level comparisons. Across more than 30,000 participants from ABIDE (autism spectrum disorder), ADNI (Alzheimer's disease), and the UK Biobank (depression), IC-GNN achieves the highest performance compared with 15 baseline methods. We further demonstrate that these individualized communities are reproducible within scans and across sessions, and carry predictive power on their own. Furthermore, IC-GNN identifies disease-related shifts in community organization that align with established literature, highlighting a promising approach for modeling individualized community structure for disease prediction and biomarker discovery.
Banks now train foundation models on ordered sequences of customer events such as transactions, app sessions, and authorizations. Deployed, such a model predicts forward in time, often for customers and merchants absent from training, yet each is evaluated on private data, so no two can be compared. Public benchmarks for such sequences rarely test a period the model has not seen; recommendation studies usually hold out each user's last events instead. We present INTENT (INTEraction, Novelty, and Transfer evaluation for next-action prediction), a public benchmark that scores a model in seven cells defined by whether the user, the period, and the target item appeared in training. An audit of 59 public dataset families retains five datasets; the most common defect is the lack of a calendar timestamp that orders events. No date lies within the time span of all five datasets, so each is split at its own cutoff, chosen by one rule fixed before any model is evaluated. Summed over the datasets, every cell holds at least 111,769 users. In reference results the cell determines the best model: a transformer leads on seen items after the cutoff but scores zero on new items, a variant that embeds item metadata helps mainly on MicroLens, and repeating the user's last item leads on new items in four of five datasets. This parameter-free control also shows that on three datasets most of the score lost after the cutoff reflects changed behavior rather than the model. The pipeline, the cutoff rule, and the per-cell user counts are publicly available.