Record- and Token-Level Data Provenance for AI Training Datasets
Oblame tackles a key unlearning gap: when contributors demand removal, trainers need exact “forget sets” at record and token granularity, not coarse dataset-level deletions. It propagates author identity through processing pipelines and resolves revocation requests into deterministic forget sets, cutting dataset-level over-deletion dramatically on Wikipedia while keeping throughput overhead modest. Tests on a 1.7B model show provenance-based forget sets improve unlearning by 42% versus random baselines.
Solving the TopSpin puzzle via cycle intersection (CayleyR)
CayleyR turns permutation puzzles like TopSpin(n,k) into a Cayley-graph search problem by generating cycles in the symmetric group and using cycle intersections to find connecting paths between states. The method uses iterative bidirectional search with a distance-guided bridge strategy when intersections are missing, and it supports optional Vulkan GPU acceleration for faster state handling. It’s packaged as an R tool with a C++ hash-indexed state store and is publicly available on CRAN.
Networked Intelligence: Active Shared Context Graphs for Human-AI Team Science
Mycelium argues that many breakthroughs can’t be achieved by scaling a single reasoning agent, because teams solve problems by sharing priors, tacit knowledge, and domain intuition across contexts. The system builds an active shared workspace that connects researchers and AI agents, routing hypotheses and observations to the right person or agent for the next action. Evaluated on a real biological multi-omics discovery workflow, shared context helped convert local findings into cross-expert mechanistic constraints and shaped better experimental design.
Learning Safe Agent Behaviour from Human Preferences and Justifications via World Models (DROPJ)
DROPJ targets safe RL in unknown environments without a usable reward function by learning a world model from real trajectories and then collecting preference feedback from humans inside that simulator. Instead of just choosing segments, humans also provide justifications, which train a reward model that is used with model predictive control for deployment. Real-user experiments suggest generating informative simulated trajectories can cut training compute and that preference-based feedback with safety justifications improves both safety and user-prioritized constraints during deployment.
Self-Improvements in Modern Agentic Systems: A Survey
This survey reframes self-improving agents as controllable evolutionary systems that update either foundation models or the operational scaffold around them (prompts, memory, tools, and control logic). It formalizes self-improvement as a self-induced update operator and organizes prior work by update targets and the signals that trigger change. The result is a system-level map of where current methods shine, how they’re evaluated, and what open problems remain for safe and reliable agent evolution.
TikTok is testing an AI likeness detection tool
TikTok is rolling out an opt-in tool that scans for AI likenesses, letting creators report suspected AI impersonation to the company. The test runs with a limited set of US creators and requires identity verification via a third-party provider (Jumio), including a real-time selfie and ID check. The move follows YouTube’s earlier rollout of similar AI likeness detection tooling.
SPINE: Bridging the Cyber-Physical Gap with Agentic AI
SPINE is an agentic framework aimed at the “spinal cord” of embodied AI – the painful calibration and debugging step between a robot’s foundation model brain and real-world hands-on operation. It uses two coordinated multi-agent workflows: a profile builder to create robot-specific context and a debugger that cycles through diagnosis, repair, and validation until teleoperation works. Across multiple DOBOT and AgileX debugging scenarios, a robotics novice using SPINE reached higher operational success and reduced time-to-teleoperation versus a baseline human using the same reference materials.
Oracle Agent Memory as an Enterprise Memory Substrate for Long-Horizon AI Agents
This work proposes “Oracle Agent Memory” as a database-native memory layer for long-horizon agents, focusing on lifecycle management like ingestion, consolidation, retrieval, and scoped revision or removal. Instead of treating memory as retrieval over flat chat history, it uses an active memory core with explicit scope control across users, agents, and threads, plus memory-centric evaluation metrics like evidence retrieval and token efficiency. Reported LongMemEval results reach 93.8% accuracy with about 10.7x fewer tokens than flat-history baselines.
Interventional Grounding Audits: Black-Box Premise-Dependency Tests for LLM Chain-of-Thought via Predicate Substitution
This paper introduces a step-level audit for chain-of-thought: intervene on a single premise by substituting a target predicate, rerun the model, and check whether each reasoning step’s conclusion changes. On ProntoQA, the method detects proof-tree dependencies far better than a self-consistency baseline, including a notable “right answer, wrong reasoning” blind spot where steps stay insensitive under consistent substitution. The authors publish audit certificates, raw outputs, and reproduction scripts to support verification.
Improving Molecular Property Prediction in Small Language Models Using Graph-based Tools
The approach improves small language models’ molecular property predictions from SMILES by injecting graph-derived context at inference time. A GNN provides predictive hints plus an instance-specific explanatory subgraph (including subgraph SMILES and a rationale), turning a mostly text-only model into a tool-augmented reasoner. Experiments on MUTAG and Tox21 show prompt enrichment can yield large relative gains, with additional intervention tests validating that extracted motifs are functionally relevant.
Apple’s plot to crush OpenAI
The Verge explains Apple’s lawsuit against OpenAI as more than a technical dispute, framing it as a strategic attempt to shape competition while Apple scales its own AI products. The coverage walks through what Apple is seeking, why it’s choosing a public fight, and how this timing intersects with Apple’s broader Siri and AI rollout efforts. The episode emphasizes that the allegations are being interpreted through the lens of how the industry typically operates, not just the words of the complaint.
The risk of weather data sabotage is rising
This report spotlights a growing threat model: because weather forecasts underpin decisions across aviation, power grids, and agriculture, tampering with weather data can create real-world harm. It connects AI-era data pipelines and forecasting dependencies to the plausibility and impact of sabotage, warning that attackers don’t need to “hack the forecast” directly to cause damaging downstream effects. The core message is that resilience and provenance for weather data should be treated as a security priority, not just an operational detail.
Fine-tune video and image models at scale with NVIDIA NeMo Automodel and 🤗 Diffusers
NVIDIA’s pipeline combines NeMo Automodel with Diffusers to make fine-tuning image and video generation models more repeatable and scalable. The focus is on practical training workflows – automating parts of the tuning process while leveraging Diffusers tooling – to reduce manual engineering when pushing models to new domains. The result is a more production-aligned path for adapting generative models to specific datasets and video/image tasks.