This paper introduces a recursive formula to efficiently compute the stochastic complexity (shortest code length) of vectors with cluster structure under the Normalized Maximum Likelihood (NML) model. The proposed approach reduces the time complexity from polynomial to linear, benefiting data clustering based on the Minimum Description Length (MDL) principle.

Researchers propose RW-LoRA, a decentralized LoRA fine-tuning method using a single model that traverses the network sequentially via random walks. This approach removes global synchronization and aggregation errors, cutting communication and computational costs. It shows competitive performance with less overhead than gossip-based LoRA across NLP tasks and graph topologies.

This paper investigates how fine-tuning erodes in-context learning (ICL) in large language models and challenges the reliability of attention-based diagnostics for preservation. The authors formalize attention-level In-Context Sensitivity (ICS) and behavioural ICL-GAP metrics. On Llama-2-7B, an ICS-maximizing regularizer nearly saturates ICS, but ICL-GAP stays near zero while MMLU accuracy drops, revealing a Goodhart dissociation. Mechanistic analysis shows attention becomes sharp and near-disjoint across prefixes yet routes to formatting and demonstration tokens rather than labels. Random-label and constructive experiments confirm the dissociation. The key lesson: attention-level ICL proxies should be used as training targets only after validation against behavioural gaps.

ReNFT proposes a novel method to repair mode collapse in reward post-training through internal probability-mass recalibration. By leveraging unconditional probes, paired counterfactual proposals, and adaptive flipping guards, it preserves 98.9%-99.0% of the reward while increasing diversity metrics by ~55%-59%, offering a complementary alternative to external interventions.

Materials property prediction in low-data settings remains a challenge, as high-accuracy models often require crystal structures that are unavailable early on. The DISTAL framework combines self-supervised compositional pretraining with structure-aware knowledge distillation from a pretrained ALIGNN teacher, enabling accurate predictions without structural inputs at inference. Across 39 benchmark tasks, the best configuration improves on 37, demonstrating complementary priors for composition-only prediction.

This paper investigates the loss of convergence when Relational Concept Analysis (RCA) is applied to AOC-posets instead of full concept lattices. The authors explain why convergence is not guaranteed in general, identify conditions for convergence, and propose both data transformations and a convergent variant that preserves the AOC-poset structure by never removing relational attributes.

Proposes REAL-Q, a novel post-training quantization (PTQ) method that directly optimizes an end-to-end aligned surrogate loss using fine-grained dynamic block-wise gradient descent and a sliding window mechanism, reducing end-to-end KL divergence by up to ~49% on LLaMA-3.1 and Qwen3 models compared to state-of-the-art methods.

This research presents TPGC, a prompt initialization method for graph pre-trained models that combines task and structural priors. It achieves better performance in few-shot settings across six benchmarks with fewer tunable parameters and lower runtime.

A study examines the relationship between prediction error and causal estimator performance in partially linear models via Monte Carlo simulations, comparing OLS, GAM, XGBoost, and DML-XGBoost. It finds prediction error doesn't consistently reflect causal bias and suggests caution in using it as a quality measure.

EULER is a multi-agent system that uses 'bridges'—transfers of problems between mathematical subfields—as its unit of search. It runs direct, adjacent-domain, and distant-domain routes in competition around a fixed conjecture, with six ordered stress tests rejecting invalid bridges before expensive search. Evaluated on 120 recent conjectures, EULER produced 10 proofs, 3 refutations, and 45 scoped partial results. Ablation studies showed bridge-specific stress tests cut incorrect conclusions from 9 to 3, and combining bridge material with target-native operations yielded a positive interaction of +4.2 resolved tasks. Domain distance did not reliably predict success; executable operation gain and valid return did.

UI-Venus-2 is a general-purpose foundation GUI agent for mobile, web, and desktop environments, operating via a unified closed-loop reasoning-action framework. It scales three dimensions: environments (over 170 multilingual mobile apps and native desktop OS), tasks (function-grounded instruction generation via deep-research pipeline), and verification (trace- and sample-level evaluators with visual keypoints and multi-model voting for reliable RL signals). Safety-aware mechanisms ensure controlled execution. This open-source model advances toward generalizable, verifiable, self-reflective agents for real-world use.

Computer science papers often rely on diagrams such as architecture drawings, flowcharts, and schematics that can convey more information than text. However, there is no public dataset pairing these figures with captions, context, questions, answers, and step-by-step reasoning, which is needed for training vision-language models. To address this, we introduce SCAFFOLD, a large-scale structured dataset of research figures with diagram QA and Chain-of-Thought reasoning traces. Using layout detection and PDF parsing, we extracted figures from arXiv papers and generated questions via AI assistance. The dataset is available in three sizes: SCAFFOLD-157K (3,058 papers, 29,887 figures, 157,387 pairs), SCAFFOLD-37K (36,797 pairs), and SCAFFOLD-12K (12,000 pairs). Baseline experiments were conduct…

OpenAgentFlow proposes a control-plane/action-plane architecture to enforce safety at the action-commit boundary. It normalizes GUI actions, API calls, tool calls, and LLM-generated invocations into a unified AgentEvent stream, routed through a shared pre-execution Policy Enforcement Point, enabling system-wide governance for heterogeneous AI agent fleets. Experiments on Android show 94.0% accuracy and 95.3% attack block rate on an action-event benchmark.

A new paper investigates LLMs' ability to track state across long tasks by having them compute MD5 step-by-step, involving 196 dependent tool calls over 64 rounds. GPT-OSS-120b succeeded in most completed runs, and findings suggest that preserving the model's reasoning context and using voting on a thinking-enabled worker can improve performance.

Financial scams targeting older adults increasingly occur through multi-turn conversations via text and voice channels. Risk signals emerge incrementally, requiring models that update risk estimates continuously under resource constraints. A new study proposes a cumulative turn-based framework, constructs a multi-turn dialogue dataset for investment, charity, and tech support scams, and fine-tunes four small language models. Phi-4 and LLaMA-3.2 show stronger turn-aware risk estimation relative to parameter scale, suggesting compact models are promising for on-device fraud protection.

This paper tackles a finite-horizon multi-item capacitated lot-sizing problem where demand quantities are deterministic but arrival periods are stochastic. Each demand occurs once within a known time window and must be fulfilled by its deadline. The model makes production and allocation decisions per demand, capturing capacity competition, demand-specific backlog, and allocation-dependent inventory. The stochastic problem is cast as a discrete-time Markov decision process (DTMDP) with explicit state space, actions, transitions, and cost. Comparing with deterministic counterparts (using most likely arrival times) shows stochastic timing vastly increases state/transition counts, solution time, and memory. A genetic algorithm (GA) is proposed that searches feasible state-feedback policies an…

This paper introduces I-CARE, a methodology for systematically studying interference in generative machine unlearning, where interference refers to the unintended degradation of semantically related concepts that should be retained. Rather than proposing a new benchmark or algorithm, I-CARE provides formal definitions for tasks, metrics, and reporting templates to enable reproducible analysis. The authors demonstrate its applicability through experiments with state-of-the-art algorithms and datasets, and provide open-source code and a web interface for exploration.

A study on how state serialization structure affects textual world model learning. HyperWorld compares raw observations with three symbolic serializations and finds that entity-centered hyperedge units yield the clearest gains for medium-small models and under distribution shift, improving feasibility detection and effect prediction, and achieving higher success in downstream planning.

The Australian government has finally proposed changes to the Privacy Act, five years after first discussions. The article highlights the urgent need for reform, citing the popularity of 'pervert glasses' at Kmart, the proliferation of facial recognition, and microtargeting ads clogging algorithms. Polls show 93% of Australians value personal information protection, and 87% are more concerned than five years ago. However, individual responsibility is impractical in the AI era, and tick-a-box consent is a broken model. Australians rank protecting personal information as their top priority for AI regulation.

Databricks has released the Big Book of AgentOps, a practical guide for building, evaluating, governing, and improving AI agents in production. The book covers agent architectures, deployment patterns, the AgentOps project lifecycle, applying DevOps principles, and operationalization steps, with insights into common pitfalls and best practices.

Botika builds agentic e-commerce tools for fashion brands and runs its entire AI stack on Modal: a 100-terabyte data pipeline, custom foundation-model research, and about 15 production inference models. CEO Eran Dagan says Modal removed the cloud-engineering overhead that once slowed them down and lets the team scale traffic dramatically without operational burden.

Apple ML researchers introduce REFACTOR-VLA, a wake/sleep system that lets vision-language-action models build reusable libraries of typed motor programs without supervision. Sleep-phase clustering uses a behavioral-equivalence kernel in a learned latent world model; wake-phase synthesis produces Hindley–Milner-inspired typed lambda terms that a rectified-flow decoder turns into actions. On full LIBERO, scaling the world model from 188M to 430M parameters hurt all four suites, while adding InfoNCE during warmup sharply improved skill-clustering NMI, up to 0.915.

Modal is opening a London office as part of a larger European investment. The company already has a Stockholm engineering team of about 20 people and now expands to London to build GTM teams, with plans to grow engineering and other functions later. European AI companies like Black Forest Labs and Legora, as well as global platforms like DoorDash, are using Modal.

Simon Willison tests Anthropic's new Claude Fable 5.1 model on his informal 'pelican riding a bicycle' benchmark across different reasoning effort levels. Higher effort levels dramatically improve the quality of the generated SVG, with the max setting producing the best pelican yet, at a cost of $3.30 and nearly 14 minutes. He then animates the image by piping it back through the model at high effort. The article also notes Fable 5.1's impressive 52.6% score on the new Terminal-Bench-Science benchmark.

Google is reportedly seeking licensing deals with major Hollywood studios to train its AI models on copyrighted content, offering hundreds of millions of dollars. While these deals could boost studios financially, they pose risks including public backlash and worker displacement. Studios may gain short-term benefits but could alienate audiences and employees.

Anthropic's new AI models, Fable 5.1 and Mythos 5.1, aim to address customer criticisms about price, data retention, and overzealous safeguards. Fable 5.1 offers stronger performance at a lower cost, with early users praising its coding abilities and nuance detection. However, Mythos 5.1 is limited to Project Glasswing participants.

The FDA developed HALO (Harmonized AI and Lifecycle Operations for Data), an enterprise data platform on Databricks on AWS GovCloud with Unity Catalog, providing a unified and governed foundation for AI and analytics. The platform supports over 6,000 users, improving SQL query speeds by 30%, cutting compute costs by 20%, and reducing data provisioning/sharing time by 75%. The FDA achieved zero-downtime migration of 5,000+ users and 8,000+ jobs/pipelines, refactoring over 1,000 pipelines and 4,000 notebooks. Key milestones included FedRAMP High authorization, AWS GovCloud migration, and Unity Catalog adoption.

At CrowdStrike's Fal.Con 2026, NVIDIA CEO Jensen Huang and CrowdStrike CEO George Kurtz unveiled SafeMind, an agentic cybersecurity system that uses NVIDIA Nemotron models and CrowdStrike's threat data to enable a continuous coevolution of offensive and defensive AI, aiming to defend against increasingly automated attacks.

Relaticle is an open-source CRM featuring approval-gated AI writing, so AI-generated content requires human review before publication.

This is a discussion about Claude Fable 5.1, described as Claude's most advanced model for coding and knowledge work, with a link for further details.