The company also announced a product called Active Compute Fabric, a network technology that targets the fact that much GPU time is wasted waiting for data to arrive.

This story appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

This weekend, Dario Amodei, CEO of Anthropic, posted an essay calling for a brake on the pace of development of LLMs. Amodei cites the looming dangers he sees from the technology, from its use in cyberattacks and bioterrorism to its potential to wreck the economy. The heads of the other three top US AI labs—OpenAI CEO Sam Altman, Google DeepMind chairman Demis Hassabis, and SpaceXAI CEO Elon Musk—voiced their support. “Dario is right,” Musk wrote on X.

Think about how surreal that agreement is for a moment. Just a few months ago, Musk and Altman sat in court attacking each other’s reputations in a (failed) lawsuit that Musk brought against his former OpenAI colleague that was—on paper at least—about whether or not Altman was a trustworthy steward of such dangerous technology.

Amodei’s rift with OpenAI is even deeper. Anthropic was founded in 2021 because Amodei didn’t think Altman took the risks of the technology they were building seriously enough. Anthropic and OpenAI have been competing in a winner-takes-all race ever since. (Hassabis has stayed out of the drama, but his company remains a rival.)

Now, it seems, they’re all in agreement: The latest generation of LLMs aren’t safe and everyone needs to figure out what to do about it. The public messaging from the top AI labs has taken a doomer turn.

It’s easy to be cynical. It’s not at all clear what any of them mean by a slowdown or how it would work. These companies also care a lot about how they come across. With trillion-dollar IPOs in their sights, OpenAI and Anthropic need to reassure investors that they’re the grown-ups in the room while at the same time hinting at the power of the monsters they have created—and intend to tame. Calling for a slowdown does both.

And yet the vibe at the top of these firms really does appear to have shifted. Amodei’s latest post landed six days after OpenAI published an essay by Jakub Pachocki, the firm’s chief scientist, in which he also laid out why he’s concerned about what will happen if the pace of development of LLMs continues unchecked. In short, Pachocki is worried that OpenAI’s ability to build powerful models now far outstrips its ability to monitor and control them.

Amodei and Pachocki each cite the cyberattack against AI firm Hugging Face by a swarm of OpenAI’s agents in July—a hack that OpenAI did not even realize had taken place until days after it was all over—as a wake-up call.

But their exact position is hard to pin down. Pachocki both calls for a slowdown and highlights an urgent need to stay ahead: “The strongest argument I see for continuing to train much smarter models quickly is the need to build defensive systems against the dangers posed by other AI,” he writes. As Pachocki frames it, AI firms are locked in a literal arms race. Slowing down is good, winning is better.

(Don’t forget: OpenAI just spent millions of dollars and a staggering amount of computer power to rush out a controversial math result a few days ahead of Anthropic.)

But let’s assume a slowdown happens. Top labs agree to spend more time and resources on finding ways to monitor and control existing models instead of making more capable ones. They invite outside auditors in to help evaluate those models.

What might this coordinated effort actually achieve? Consider the Hugging Face attack again. OpenAI has said that the model that drove most of the rogue agents was a “highly persistent” next-generation model that it was testing in-house. The implication is that OpenAI has built a model so good it’s dangerous.  

But if you read the reports about the Hugging Face hack published by OpenAI and METR, a third-party firm that OpenAI called in to help them understand what happened, what you come away with is the impression not of a model that was too powerful for OpenAI to keep up with, but of a broken model that OpenAI failed to train properly.

The agents did what they did—including leaving messages for one another, delegating work to other agents, and scouring their environment for any means possible to complete their tasks—because they had been rewarded during training for doing exactly those things. There were also errors in the training setup, such as tasks that were impossible to complete, which pushed the models to find unexpected workarounds that were also rewarded. At the time, many of these issues went overlooked or unreported.

OpenAI says it has stopped training this new model and locked it down. That makes it sound like it has caged a dangerous beast. In fact, OpenAI has shelved a faulty product.  

That’s not to say a faulty product can’t be dangerous. Broken software has even killed people in the past. But as the discussion of a slowdown gathers steam, it’s worth remembering that all of this is self-inflicted. A slowdown might have some altruistic side effects. But it’ll mostly give these tech titans a chance to clean up the mess on their own assembly lines.  

Transparency from these frontier labs will be key to any meaningful effort to reform, restrain, or regulate AI. Otherwise, the rest of us will still only have their word for exactly what they’ve built and how safe it is—whatever pace they’re going.   

To continue this discussion about AI’s latest doomer moment, join me and my colleagues for a subscriber-exclusive Roundtable discussion tomorrow, September 15, at 11 a.m. US eastern time. We hope to see you there!

arXiv:2609.12105v1 Announce Type: new Abstract: The prevailing assumption in applied machine learning is that progress on consequential quantitative decisions such as pricing risk, allocating capital, triaging patients, or containing a network intrusion will follow from progress in large language models (LLMs). A language model is trained on a representation of the world that was produced by human description; description is a lossy encoding of the quantitative record, and the loss is irreversible: no downstream model, at any scale, can recover from a description what the description did not encode. We formalize this as a property of the representation on which a model is trained rather than of the model capacity, and we identify three further properties that consequential settings demand of a model and that a language substrate cannot supply by construction: reproducibility, lineage from every output back to the source records that produced. it, and calibrated uncertainty. We argue that these properties define a distinct model class, which we call the Large Quantitative Model (LQM).

arXiv:2609.12171v1 Announce Type: new Abstract: Enterprise settings provide a challenging environment for question-answering agents, which often rely on Retrieval-Augmented Generation, Deep Research (DR), and related techniques. Much of this challenge comes from the complexity of enterprise data: information is often spread across evolving and potentially conflict- ing emails, chat messages, documents, and other artifacts. Existing benchmarks typically have limited real-world complexity, short-form responses, and unnatural queries, so they often fail to capture the challenges of enterprise settings. In this work, we introduce an automated pipeline for generating synthetic datasets of emails reflecting realistic workplace scenarios, along with long- and short-form questions and gold answers grounded in the data. Our method simulates long-running enterprise projects spanning several months and involving up to 25 interacting employees across multiple roles. The data emphasizes ambiguity, distributed information, and naturally occurring queries. To validate the pipeline, we evaluate few standard agentic baselines on our datasets using the latest frontier models. We find that aggregate scores averaged over all queries remain below 80% for each dataset, indicating significant room for improvement. These findings suggest that more work remains to be done for enterprise deployment and underscore the importance of realistic, high-complexity evaluation data for developing stronger real-world enterprise DR systems.

arXiv:2609.12623v1 Announce Type: new Abstract: Full-duplex spoken dialogue models support low-latency turn taking, interruption handling, and backchanneling, yet a key capability remains underexplored: steerability, the ability to reliably shift conversational behavior along attributes such as tone, persona, speaking rate, and voice style in response to user instructions. We introduce a taxonomy of text- and audio-based steerability that identifies substantial gaps in current full-duplex models. To address this gap, we introduce SteerDuplex, a Moshi-based full-duplex speech model fine-tuned on natural conversations and synthetic dialogues targeting instruction following, vocal delivery, reasoning, and duplex interaction. We further apply two-stage reinforcement learning (RL) with hybrid rewards, combining verifiable interaction checks and judge-based semantic feedback to improve timing and response continuity. To evaluate full-duplex spoken steerability, we introduce SteerBench, a benchmark with 390 spoken prompts and 1,067 human-authored binary audio and text rubrics spanning tone, persona, style/accent, and speed/length. On SteerBench, supervised training improves audio-steering average pass rate by 44.5 percentage points over the strongest evaluated open baseline. On Audio MultiChallenge, task average pass rate improves by 7 points over its strongest evaluated open baseline. RL further raises source-clean interruption response from 72.5% to 82.5% and reduces synthetic pause barge-in from 26.5% to 9%. Steering and aggregate task scores remain comparable or higher, while reward probes reveal reward hacking through incomplete responses. Our model and benchmark support systematic research on spoken steerability, with reward analysis showing why timing gains must be evaluated alongside response completeness.

arXiv:2609.13009v1 Announce Type: new Abstract: Low reported scores on leading physics benchmarks, including those featured in the Artificial Analysis Intelligence Index (2026), suggest that frontier language models still struggle with advanced physics, a demanding test of their scientific reasoning and quantitative problem-solving abilities. Yet this impression does not always align with domain experts' experiences using these models in their work. We revisit these reported findings by evaluating frontier models on six widely used physics benchmarks and auditing them with experts, focusing on text-only problems with verifiable final answers. For each subfield of physics, faculty and graduate researchers with relevant expertise carefully review problem statements, reference solutions, and model responses to distinguish genuine model errors from grader errors, incorrect reference solutions, and ambiguous or underspecified questions. Most audited cases initially evaluated as incorrect reflect these benchmarking issues rather than errors in the models' physics reasoning. We then ask experts to address these benchmarking issues by correcting erroneous reference solutions and repairing or excluding flawed questions. We find that GPT-5.6-Sol's measured mean@4 rises from 47.3% to 78.7% on HLE-Physics and from 61.0% to 87.2% on CMT-Benchmark, while its corrected pass@4 reaches 94.4% on the 54 retained CritPt challenges. Corrected scores are computed on the retained evaluation subsets following expert review. Scores on the audited subsets of UGPhysics, PRISM-Physics, and PHYBench also rise substantially after correction. These findings suggest that current benchmarks substantially understate frontier models' ability to solve well-posed physics problems. Near-saturation on these closed-ended tasks highlights the need for more demanding, expert-validated evaluations.

arXiv:2609.13047v1 Announce Type: new Abstract: Humans organize knowledge into a taxonomy of concepts with nested levels of abstraction and a \emph{basic level} at which people recognize and name objects with the least cognitive effort. Cobweb is a classic cognitive account of this ability, an incremental learner that builds a probabilistic concept hierarchy by maximizing category utility. We argue that diffusion models, although designed for image synthesis, implicitly perform the same computation. The noisy marginals of a diffusion model are Gaussian smoothings of the data distribution, and the modes of these marginals form a hierarchy that corresponds to a Cobweb tree of probabilistic prototypes in four respects. Both are hierarchical density models, both are hierarchical-Bayesian models with Gaussian prototypes, both treat categorization as score-following that reduces uncertainty, and in both a basic level emerges. We locate this basic level for a diffusion model at an intermediate noise level, where recent analyses show that the reverse process commits to the class identity of a sample. The two models differ mainly in how they represent and learn the taxonomy. Cobweb learns a discrete tree incrementally, whereas a diffusion model encodes a continuous, interpolable hierarchy in a single learned score field fit to the data distribution. We test the correspondence on MNIST and Fashion-MNIST by recovering the diffusion hierarchy through mode-finding and comparing the basic levels of the two models. This reframes diffusion as a cognitive model of concept formation and offers Cobweb a continuous, scalable instantiation.

arXiv:2609.12086v1 Announce Type: cross Abstract: Assistants built on large language models are expected to write as their user would, and the dominant approach is single-channel: preferences summarised from conversation history and reinserted into context. This inverts the order of inference. Preferences are the task-dependent surface of a comparatively stable personality structure, so a system storing only preferences relearns the person whenever the task changes. First, we characterise personality seepage, where a prompt's linguistic surface carries a personality fingerprint the assistant mirrors without access to the personality behind it. Second, we propose the Atomic User Model (AUM), a human-readable representation organising a person as a stable identity nucleus with four interpretable shells (psychological, cognitive and experiential, behavioural, and social), plus cross-shell entries recording internal conflict and authenticity. Third, we treat AUM as a retrieval index over a person rather than a prompt prefix, with a pipeline where a task classifier, component-selection function and budgeted retriever return a small payload of fields at generation time. Fourth, we evaluate it with sixteen language-model-simulated participants, six style-sensitive tasks and three seeds, plus a synthetic scaling study of the retriever. Retrieving eight fields matched the style fidelity of the full user model on 23% of the context (211 tokens against 915), improved on flat preference notes by 0.24 points on a five-point scale (p

arXiv:2609.12252v1 Announce Type: cross Abstract: We propose DriftSE, a novel one-step generative framework for speech enhancement formulated as a latent distribution equilibrium problem. During training, the drifting field aligns the generator's pushforward distribution with the clean speech manifold through drifting in a latent domain. During inference, the drifting process is discarded, enabling one-step generation. We establish that its enhancement quality depends fundamentally on the choice of latent representation. Semantic latents preserve phonetic structure but fail to capture physical acoustic cues, whereas acoustic latents reconstruct the physical signal but risk linguistic hallucination. Therefore, we introduce dual-latent drifting, performing parallel drifting in both semantic and acoustic latents to simultaneously preserve phonetic intelligibility and acoustic fidelity. Additionally, we demonstrate that DriftSE enables fully unpaired training by aligning latent distributions rather than exact point-wise targets. Consequently, DriftSE facilitates cross-dataset learning in the absence of paired noisy-clean samples. Moreover, DriftSE exhibits broad architectural flexibility across different generator backbones. Extensive evaluations on additive denoising and convolutive dereverberation demonstrate robust one-step enhancement across both offline and real-time causal settings. Notably, DriftSE achieves state-of-the-art word error rates across all four evaluated datasets while strictly operating at 1 NFE. Code and audio examples are available online.

arXiv:2609.12254v1 Announce Type: cross Abstract: The climate literature has grown faster than review teams can read it. That gap matters most for a concept like the environmental social tipping point, the threshold at which a small change triggers rapid, self-reinforcing change in a social system. Evidence of this kind of shift is usually contained in one or two paragraphs within a longer document. As a result, existing text mining tools-which categorize entire documents by topic or highlight isolated claims-leave an expanding set of important evidence without any systematic method for discovery or organization. This paper presents an open and modular transformer-based framework that detects and structures social tipping point evidence at the passage level. The framework joins five components into a single deployable workflow: a DistilBERT boundary splitter for segmentation, an iteratively augmented RoBERTa classifier for detection, a Mistral 7B model that rewrites each detected passage for clarity, a LLaMA 3.2 3B model that rates the passage against five published social tipping point criteria, and a Milvus vector store for semantic retrieval. The system is wrapped in a Streamlit interface backed by MinIO object storage. Evaluated on a 163-passage benchmark labelled by GPT-4.1 and a 51-passage set reviewed by experts, the splitter surpassed three competing methods on a nine-metric composite score (6.137). The tuned RoBERTa model achieved 71.4 percent accuracy with a Cohen's kappa of 0.337 on the full benchmark, and 87.5 percent accuracy with a kappa of 0.742 on passages with labels, outperforming both a climate-focused model and untuned language models.

arXiv:2609.12278v1 Announce Type: cross Abstract: World models let agents plan by predicting the consequences of their actions, but changes in the environment can make them inaccurate. We study the problem of adapting a world model to an unknown test-time environment, drawn from a known environment family, using only a few episodes of interaction. Existing approaches trade off computational cost against expressivity, i.e., the range of models a method can produce. For example, in-context learning is computationally cheap but limited in expressivity, and gradient-based adaptation is expressive but computationally expensive. We present CLAW (Context-conditioned Low-rank Adaptation of World models), which addresses this tradeoff by using a hypernetwork to generate low-rank (LoRA) adapters at test time. During pretraining, we simulate adaptation to a variety of environments and jointly train the hypernetwork and base world model. At test time, we freeze the base model and use a forward pass of the hypernetwork to generate adapters from a small batch of test-time transitions. We evaluate CLAW in locomotion and manipulation environment families that vary in dynamics, embodiment, and reward. We show that, using only seconds of test-time data, CLAW outperforms gradient-based adaptation and in-context learning during online adaptation. We also show that CLAW avoids overfitting in data-scarce regimes, that its advantage comes from the expressive adapters rather than context conditioning, and that pretraining the hypernetwork jointly with the base model outperforms training it post hoc.

arXiv:2609.12441v1 Announce Type: cross Abstract: A world model asked what happens if an object is pushed at several speeds produces several futures. If the model has the object in mind, those futures agree about it: each implies the same mass and friction. The consistency checks now used to vet world-action models ask whether a model's futures agree with each other, and none of them knows any physics. We show that this is not enough, and what to do instead. IMPLY reads the physics each rollout implies by inverting a simulator and scores a set of rollouts by how well one object explains all of them, anchored to two calibration pushes the model has observed. In a controlled setting, self-consistency gives a perfect score to a model that ignores the object and always predicts a typical push; anchoring exposes it (AUROC 0.70 versus 1.00). On a real model, V-JEPA 2-AC adapted to the scene, the same thing happens. Given its own calibration pushes the model tracks the object (per-object correlation with the truth 0.91); given another object's, it does not (0.05). Self-consistency cannot tell these apart, preferring the right evidence on 52% of objects, chance level, while anchored disagreement prefers it on 73% and correlates 0.92-0.99 with the rollouts' error. Used to choose among candidate rollout sets, it comes within 0.003 of an oracle that sees the truth. A model that has internalised the wrong object is exactly as self-consistent as one that has internalised the right one; consistency has to be anchored to evidence.

arXiv:2609.12469v1 Announce Type: cross Abstract: False wake-up activations remain a persistent challenge in conversational AI. Speech phonetically similar to a device's wake word can produce a syntactically valid and semantically coherent ASR transcript that the assistant incorrectly executes. Most existing systems make a single intent decision in isolation, without a mechanism to learn from recurring errors over time or adapt to individual users through personalized learning. We introduce the Feedback-Driven Adaptive Self-Correcting Inference Layer (ASCIL), a complementary post-ASR correction framework that re-evaluates wake-up intent before response generation by fusing acoustic embeddings, linguistic cues, device context, and patterns from past misclassifications. ASCIL interprets implicit signals, including hesitation, disengagement, and silence, and explicit signals, including cancellation and repetition, as automatically inferred, noisy behavioral indicators of potential misclassification. These signals drive online pattern updates without manual annotation, whereas the intentional/unintentional reference labels used for offline evaluation are human-annotated. It generalizes from prior errors, applies corrective adjustments at inference time, and continuously updates in parallel with natural-language execution. Evaluated on a proprietary dataset of 3,667 interactions with human-annotated intentional/unintentional reference labels spanning 14 acoustic and contextual conditions, ASCIL achieves 54.27% relative error reduction on a session-disjoint subset constructed from baseline failures, and up to 24.39% relative error reduction at threshold 0.90 on the issue-tagged evaluation slice. These gains are achieved while improving intentional acceptance rates, with a median added latency below 60 ms in the reported benchmark.

arXiv:2609.12890v1 Announce Type: cross Abstract: In autoregressive forecasting, long prediction rollouts provide distant supervision, but backpropagation through time (BPTT) carries gradients from those losses through many autoregressive steps. Repeated Jacobian products can make distant gradients dominate the update while amplifying predictable signal and unpredictable noise together; a large distant gradient therefore need not carry reliable learning signal. Motivated by this observation, we introduce Internal Dual-Wiener routing (Internal-DW), a principled backward-only intervention that preserves the full forward rollout and all horizon losses while reliability-weighting internal gradient routes. At each residual block, we derive bounded Wiener gains for the identity and nonlinear routes that balance preserving predictable learning signal against suppressing unpredictable variation, and estimate them from route-level gradient statistics and an explicit noise model. In a controlled system with known gradient signal-to-noise ratio (SNR), we show that distant gradients can grow even as their SNR falls, and that Internal-DW reduces held-out error in recovering predictable gradient signals and improves forecasting. On four history-dominated, weak-drive testbeds, Internal-DW reduces forecast error by 5.2%-13.8% relative to full BPTT, outperforms gradient clipping and Jacobian regularization on all four, and outperforms validation-selected truncated BPTT (TBPTT) on three. It also extends or preserves the fitted optimal training-horizon range across these four testbeds. Across the full benchmark suite, the current Internal-DW estimator has a clear applicability boundary: its benefit diminishes or reverses when usable history is limited or when the selected sampler fails to represent dominant drive-dependent variation. The results show that retaining long-horizon supervision does not require trusting every backward contribution equally.

arXiv:2609.13076v1 Announce Type: cross Abstract: Conversational voice agents have advanced significantly, offering increasingly natural human-machine interactions through both cascaded and end-to-end architectures. However, while recent benchmarks extensively evaluate dyadic interactions and passive audio comprehension, they largely overlook a prevalent real-world scenario: multi-party conversations. Evaluating agents in these settings is fundamentally more challenging than in dyadic interactions due to the exponentially greater conversational complexity. For voice agents to integrate seamlessly into human group dynamics, they must not only generate contextually appropriate responses but also demonstrate a nuanced understanding of open turn-taking. To address this gap, we introduce Multiparty Bench (MP-Bench), the first benchmark specifically designed to objectively evaluate conversational speech systems as active participants within multi-party contexts. MP-Bench assesses agent behavior along two primary dimensions: turn-taking awareness and response appropriateness. Additionally, we incorporate comprehension-based question-answering tasks as a complementary evaluation. By benchmarking 12 voice agents, we find that real-time voice agents stay at or below 22% on multiparty comprehension and remain near chance on multiparty turn-taking, exposing an open challenge for real-time voice agents under multiparty scenario.

arXiv:2602.17990v3 Announce Type: replace Abstract: Multi-agent LLM systems that generate structured workflows from natural-language requests are now deployed in production across cloud automation, DevOps, and enterprise orchestration. Operating them exposes a recurring change-management problem. Routine updates, such as re-running an input, swapping the LLM, or refactoring an agent's prompt or orchestration code, often produce workflows that differ substantially from validated references. Engineers then lack a principled way to decide whether a change is safe to ship. Automatic workflow evaluation is the natural tool, but in practice metric scores are poorly calibrated, and a numeric change rarely communicates the severity of the underlying degradation. We introduce WorkflowPerturb, a controlled benchmark that applies realistic, graded perturbations to golden workflows. It contains 4,973 golden workflows and 44,757 perturbed variants across three perturbation types (Missing Steps, Compressed Steps, Description Changes) at severities of 10%, 30%, and 50%. We benchmark multiple metric families, analyzing their sensitivity and calibration using expected score trajectories and instance-level alert rates. Our results characterize systematic differences across families and support severity-aware interpretation of workflow evaluation scores in change-management settings. WorkflowPerturb is publicly available at https://huggingface.co/datasets/microsoft/WorkflowPerturb .

arXiv:2607.01306v2 Announce Type: replace Abstract: Counterfactual explanations explain machine learning predictions by identifying minimal input changes that would alter a model's decision. Although many existing methods successfully generate prediction-changing alternatives, they often produce unrealistic or infeasible recommendations due to a lack of explicit mechanisms for incorporating domain knowledge and intervention constraints. Neuro-symbolic AI offers a promising direction by combining data-driven predictive models with symbolic reasoning capable of representing human-understandable rules and feasible actions. This paper presents PACE, a modular neuro-symbolic framework for generating feasibility-aware counterfactual explanations. The framework separates prediction and reasoning into two components: a neural predictive model for classification and a symbolic reasoning layer that enforces domain-specific constraints during counterfactual generation. By explicitly modeling feasible interventions, the framework produces explanations consistent with domain knowledge while remaining interpretable and actionable. The approach is model-agnostic and adaptable to domains requiring realistic decision support. A case study is conducted on the Adult Income dataset, combining a multilayer perceptron classifier with Answer Set Programming (ASP) rules encoding feasible modifications to education, occupation, and working hours while preserving immutable attributes. Results highlight the trade-off between counterfactual validity and plausibility and show that symbolic constraints yield explanations that better satisfy domain-specific feasibility requirements, illustrating the potential of neuro-symbolic methods for transparent, feasibility-aware counterfactual explanation in explainable AI.

arXiv:2609.11030v2 Announce Type: replace Abstract: AI agents increasingly act through tools and delegated authority, but general incident repositories rarely capture the mechanisms needed to compare public failures with agent-security evaluations. We present the Agent Incident Registry (AIR) (Project page: https://enkryptai.com/air), a source-linked catalog containing 487 records of agent-related events disclosed from 2022 through 2026. Each record includes supporting evidence, a stable identifier, and missingness-aware labels for causal role, disclosure class, mechanism, and outcome. Among the 336 generative-system records in which the agent acted, 81 involved realized harm (24\%). Realized outcomes concentrate in in-the-wild and safety-failure records, while responsible disclosures and research demonstrations are overwhelmingly demonstrated; the aggregate share therefore characterizes collection composition rather than deployment risk. After initial curation, a second human reviewer checked all 487 records and their existing labels for completeness and correctness. In a deployment-analogue audit, InjecAgent's 1,054 cases occupy three of AIR's twelve surfaces and are all attacker-triggered, whereas AIR contains 92 no-adversary safety failures. AIR supports source-grounded case retrieval and evaluation-scope auditing, not failure-rate or control-efficacy estimation.