Scored 10 articles from 3 feeds; 10 included in digest.
Run ID: run-1789026704809
Generated: September 10, 2026 at 03:52 AM ET
Summaries: gemini-flash-latest; enrichment 10/10 succeeded
| Source | Type | Included | Scored | 28d Digest Rate | 28d Avg Score | 28d Hotlist Hit | 7d Article Age | 28d Confidence |
|---|---|---|---|---|---|---|---|---|
| OpenClaw: discovery-rank | curated | 10 | 10 | ~100% | ~0.31 | ~0% | Unknown | Low sample |
Source: OpenClaw: discovery-rank
Type: curated
Included: 10
Scored: 10
28d Digest Rate: ~100%
28d Avg Score: ~0.31
28d Hotlist Hit: ~0%
7d Article Age: Unknown
28d Confidence: Low sample
Researchers introduce OpenDiscoveryTrace, a publicly available dataset under CC BY 4.0 comprising 558 complete execution trajectories of AI scientific agents across 124 tasks in genomics, drug discovery, materials science, and scientific literature analysis. Unlike traditional benchmarks that evaluate only final outputs, OpenDiscoveryTrace records a structured nine-field trace for each step—capturing elements such as thoughts, tool calls, observations, errors, revision triggers, and confidence—across three frontier models (GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro) and four open-weight models. A pilot analysis revealed that while frontier models achieved comparable success rates (84–89%), their step-level behaviors diverged significantly, showing distinct error rates and types. The authors also establish five benchmark tasks with classical and neural baselines to support future research in AI scientist auditing, process evaluation, and governance.
Keywords: OpenDiscoveryTrace, AI scientists, autonomous agents, benchmarking, AI evaluation, model capabilities
In this paper, the authors introduce the State-Path Tool Menu to improve how tool options are selected and ordered for language model agents facing large tool libraries. Addressing the limitation of standard relevance ranking—which often omits or misorders prerequisite tools needed to generate inputs for subsequent actions—the framework treats the tool menu as an execution prior over a "state path" from request to final outcome. It employs an encoder to model tool executability and input-output dependencies, a retriever to identify entry tools, prerequisite producers, and final actions, and a reranker to place producer tools ahead of consumer tools. On ToolBench evaluations, the method increased online agent success from 0.737 to 0.898 without modifying the underlying agent, covered more complete execution chains with 32 tools than the official baseline covered with 128, and maintained performance advantages across diverse model capacities.
Keywords: Language models, Online agents, Tool menus, Model capabilities, Multi-step execution, Artificial intelligence
In the paper "Do Agents Know When They Succeed? Calibrating Agent Confidence from Internal Representations," researchers examine whether a model's internal representations can predict task success in multi-turn agentic systems. The authors propose two approaches: Latent Trajectory Dynamics (LTD), which tracks residual-stream representation changes across an interaction trajectory, and the Action Representation Probe (ARP), which predicts success based on representations at action decisions. Tested across Bash, SQL, and Python benchmarks using Qwen14B, Qwen7B, and DeepSeek6.7B models, both methods outperformed surface-level generation and sequence-based calibration baselines, functioning as zero-overhead reliability monitors that do not require prompt alterations or multi-sample rollouts.
Keywords: AI agents, internal representations, agentic workflows, confidence calibration, model capabilities, safety-critical AI
This paper investigates how language model agents can effectively execute libraries of reusable knowledge, known as skill packages, to solve long-horizon tasks. While skill packages are conventionally loaded directly into an agent's main context—an approach that can become brittle as reasoning degrades with expanding context windows—the authors propose invoking them as subagents with dedicated, fresh context windows for individual subtasks. The study finds that subagent execution outperforms standard agent-skill execution when skill packages provide clear input-output contracts and procedural instructions, though this approach incurs additional token overhead to coordinate communication between the main agent and subagents.
Keywords: Language Model Agents, Agent Skills, Subagents, Artificial Intelligence, Long-Horizon Tasks, AI Capabilities
The paper introduces XAI-Arena, an LLM-as-a-judge framework designed to provide scalable, reproducible, multidimensional, and stakeholder-sensitive evaluations of explainable AI (XAI) explanations. Developed to address the scalability and reproducibility limitations of subjective human judgment, the framework assesses explanations across dimensions including perceived simplicity, clarity, task adequacy, trust calibration, actionability, transparency, faithfulness, and overall interpretability across multiple datasets, models, and personas. Benchmarking and human validation revealed a strong positive correlation between LLM-generated and human ratings (Spearman's rho = .693, p < .001), demonstrating that LLM-based evaluations can capture systematic differences in XAI explanation quality.
Keywords: Explainable AI, XAI, Large Language Models, LLM-as-a-judge, XAI-Arena, AI Evaluation, Model Capabilities
This paper presents a probability-wave framework using a generalized behavioral intelligence (GBI) nonlocal probability-wave equation to model the collective behavior of interacting adaptive agents. Using empirical Chinese intraday stock market data, the authors found that adaptive entangled game modes account for 82–94% (89% overall) of observed decision patterns, whereas purely independent modes represent less than 5%, providing empirical support for the Liu-Chen-Ao (LCA) hypothesis regarding nonlocal entangled nerve fibers in the brain. Based on these findings, the authors advocate incorporating adaptive entangled game modules and probability-wave-based brain simulations into artificial general intelligence (AGI) architectures to address the opacity of conventional neural networks and enable more compact, efficient human-like processing units for embodied intelligence and robotics.
Keywords: Artificial General Intelligence, Adaptive Agents, Probability-Wave Framework, Trader Behavior, Generalized Behavioral Intelligence, Agent-Based Modeling
This paper evaluates the use of active learning to select laboratory experiments for scaling up critical-materials recovery, analyzing data from Pacific Northwest National Laboratory's Computer Intelligence for Critical Element Recovery and Optimization (CICERO) workflow. In retrospective benchmarks on recycled neodymium-iron-boron (NdFeB) magnet records, adaptive policies reached maximum recorded enrichment in 16 to 24 experimental wells, compared to 48 wells using nonadaptive space-filling methods. The authors also analyze trade-offs in recycled samarium-cobalt magnets and produced water from oil and gas extraction. To guide scale-up decisions, the authors propose selecting experimental batches by expected reduction in downstream Bayes risk, finding in simulations that a candidate-filtering hybrid policy reduced estimated loss compared to an existing joint search. Finally, they outline requirements for prospective testing under a shared loss and logging standard.
Keywords: Active Learning, Artificial Intelligence, Materials Recovery, Industrial Optimization, PNNL, CICERO
This paper presents Valerant, a training-free framework designed to autonomously generate persistent, navigable 3D game maps from a single image. While existing game-oriented world models typically operate within 2D visual observation spaces lacking persistent 3D geometry, Valerant adapts a pretrained action-conditioned world model into a World Action Model (WAM). By combining predictive visual rollouts with SLAM-based spatial reconstruction and exploration-driven action selection, the framework maps and instantiates explicit 3D geometry, offering a method to reduce manual labor in 3D game-map creation.
Keywords: World Action Models, Embodied AI, Valerant, World Models, Game Map Generation, Machine Learning
This paper presents a human-in-the-loop, multi-agent GeoAI system designed for Arctic maritime eco-navigation. Addressing the operational, environmental, and community risks that accompany expanding Arctic accessibility, the framework integrates physical, operational, ecological, and community factors into route planning rather than focusing solely on travel time, fuel efficiency, and navigational hazard. Specialized agents manage geospatial data preparation, multi-objective route generation, and skyline-based decision support, incorporating exposure risks to sensitive areas such as Essential Fish Habitat and seal critical habitat while leaving consequential value judgments to human operators. The authors have made the project page and code publicly available.
Keywords: GeoAI, Autonomous Agents, Arctic Navigation, Maritime Logistics, Route Optimization, Environmental Risk
This paper investigates the hypothesis that the first-order structure of physical interactions, such as gradients or Jacobians, characterizes phenomenal experience. Setting the analysis in "Gradland"—an idealized environment inhabited by neural networks with known physics and largely differentiable functions—the authors introduce two measures of Jacobian structure: effective rank and cohesion, based on Kirchhoff complexity. Through a series of worked examples, the paper demonstrates how this framework accounts for the duration of experience over hundreds of milliseconds, vivid versus obscure experiences, the perception of texture, newborn perceptual confusion, distinct versus confused ideas, the nature of learning, and the functional purpose of rich, dense experience.
Keywords: neural networks, phenomenal experience, Jacobian, Kirchhoff complexity, effective rank, Gradland