Eye on AI Weekly Research Watch

Knowing When to Stop: Bayesian Optimal Stopping for LLM Evaluations Aug 19, 2026 171 Authors: Toby D. Pilditch Authors: Toby D. Pilditch Paper: https://arxiv.org/abs/2608.14425v1

The Dynamics of Intelligence Explosions Aug 19, 2026 126 As AI increasingly accelerates its own research and development, could this create a runaway feedback loop of self-improvement? This theoretical paper mathematically analyzes the conditions required for such an "intelligence explosion," showing that truly explosive, asymptotic growth is harder to achieve than prior economics-based models suggested, while identifying an underappreciated mi

PACE-Bench: Benchmarking Physics Adaptation via Code Evolution in Dynamic Environments Aug 19, 2026 121 Self-improving AI agents are usually tested under fixed conditions, but real-world deployment demands adapting when the environment itself changes. PACE-Bench introduces 144 source-to-target adaptation challenges across six physics domains, forcing agents to iteratively rewrite working code when the underlying physics is mutated. Testing ten methods reveals that grounded, feedback-driven

Designing Compact Neural Architectures via Neuron Gating and Mixed Activation Aug 19, 2026 148 Neural Architecture Search is powerful but expensive due to discrete, combinatorial design choices. This paper proposes continuous relaxations of neuron-level and activation-level decisions, enabling fully differentiable optimization across MLPs, CNNs, RNNs, and Transformers. Three resulting methods (NAS-NG, NAS-MA, NAS-NGMA) find highly compact architectures --- including a CNN with just

Wyvern: An Agentic Framework for Generating Grounded Multimodal Reports Aug 19, 2026 137 As AI-generated content proliferates faster than anyone can verify it, ungrounded generative reports risk spreading unsupported claims. Wyvern addresses this with a multi-agent framework that produces technical reports combining text, images, and tables, all tied to supporting references, with a dedicated claims-revision stage to strengthen grounding. Human evaluators rated Wyvern's figur

SheetCompass: Hierarchical Relation Graphs for Agentic Spreadsheet Reasoning Aug 19, 2026 132 Spreadsheets pack in complex, spatial, and cross-referential structure that gets destroyed when LLMs flatten them into plain text, crippling automated analysis. SheetCompass instead builds explicit relation graphs capturing structural links within and across worksheets, paired with a memory system that keeps relevant context available to reasoning agents. This lets AI agents navigate spre

Shift Aware Transfer Learning with Adaptive Dual-Encoder Fusion for PM Forecasting in Data-Limited Environments Aug 19, 2026 164 Forecasting air pollution (PM2.5) is hard where local sensor data is scarce, and naive transfer learning from data-rich regions can actually hurt performance due to domain mismatch. This study's dual-encoder framework pretrains on U.S. monitoring data, then adaptively fuses it with limited Taiwan-specific data, letting the source model adjust rather than freeze. The adapted model beat bas

LP-NAS: Linear Programming-based Neural Architecture Search Aug 19, 2026 153 Automating neural network design (NAS) is powerful but computationally costly, and differentiable NAS methods like DARTS often converge slowly. LP-NAS reframes the architecture search as a linear programming problem, using gradient and Hessian information to compute better-informed update directions that improve generalization while preserving optimized parameters. Its two variants integr

Ensuring Safe Physical AI in Urban Mobility via Hazard-Informed Synthesized Envelopes Aug 19, 2026 122 As robots increasingly share city streets and sidewalks with people, ensuring safety can't just be handled by isolated software modules --- it needs to span the whole system. This paper proposes hazard-informed safety envelopes that connect systematic hazard analysis to real-time runtime enforcement, unifying symbolic, spatial, and dynamic representations of the environment. The framework

Twin: Playing an Unknown Game with a Test-Time Digital Twin Aug 19, 2026 125 Instead of hand-engineering a world model for each new task, this paper has a coding agent write executable world models on the fly, purely from observing an unfamiliar game. A "twin" simulation validates each hypothesis by requiring it to reproduce every past observation before acting, using mismatches to iteratively repair the model. Tested on ARC-AGI-3 games, Twin dramatically outperfo

Optimal Scheduling of Road Maintenance Jobs Considering Impact on Traffic Flows Aug 19, 2026 124 Planning road maintenance at scale requires repeatedly simulating how traffic redistributes when capacity is reduced --- a process that's normally too computationally expensive to embed directly into scheduling optimization. This paper builds data-driven surrogate models that predict equilibrium traffic flows directly from origin-destination demand, bypassing expensive equilibrium solvers

Generating Benchmark Health Data Using a Tabular Diffusion Transformer Aug 19, 2026 137 Healthcare researchers often need synthetic data spanning multiple related but heterogeneous tables, yet most generative models only handle single tables. This paper's two-stage framework first standardizes diverse tables into common statistical summaries capturing distributions and correlations, then uses a diffusion transformer to generate new synthetic statistical tables, which are rec

Universal Thermodynamic Interatomic Potentials for Crystalline Materials Aug 19, 2026 150 Discovering new materials computationally usually relies on ground-state energy calculations, ignoring temperature and pressure effects that determine real-world phase stability. This paper introduces TIP, which upgrades an interatomic potential model into a full thermodynamically consistent free energy model using automatic differentiation, letting a single calculation reveal phase trans

RecipeNet: A Hierarchical Transformer for Recipe Data Aug 19, 2026 175 Many real-world processes --- chemical synthesis, drug formulation, manufacturing --- are naturally represented as ordered, structured steps rather than flat tables, yet most machine learning models flatten this structure and lose crucial dependencies. RecipeNet addresses this with a hierarchical Transformer that separately models relationships within each step and dependencies across the

Split the Labor: Separating Evidence Interpretation from Decision Aggregation Aug 19, 2026 157 When LLMs are asked to synthesize conclusions from many sources, cramming everything into one prompt conflates two different jobs: understanding each piece of evidence and combining conclusions fairly. This paper identifies a subtle failure mode, "count-scale drift," where naive vote-counting shifts as more sources are added, and proposes separating interpretation from aggregation using a

Learning-to-Transition for Large-scale and High-Order MIMO Detection Aug 19, 2026 127 Modern wireless systems need to decode signals sent simultaneously from many antennas, a computationally brutal search problem especially at high signal orders. This paper reframes MIMO detection as a learned sequential decision process, using a Transformer to guide symbol-by-symbol search while producing reliable confidence scores for downstream error correction. A clever "hard-to-soft"

Participatory Moral AI Is Not Neutral: The Invisible Hand of Developers Aug 19, 2026 150 Democratic AI alignment approaches poll people on moral dilemmas to train policies --- but this study shows the "neutral" aggregation process is shaped by hidden developer choices. Across studies involving kidney allocation, AI worker simulation, and deepfakes of the deceased, the researchers show that which features get voted on, who votes, and how questions are worded all systematically

Handover of In-Context Learning State Across Session Boundaries Aug 19, 2026 118 When an LLM conversation exceeds context limits, restarts, or gets handed to another agent, someone must decide what information survives the transition. This paper formalizes that handover problem mathematically, distinguishing exact recovery from statistically sufficient preservation, and proposes a three-part memory record for decisions, summarized evidence, and irreplaceable raw obser

Marionette: Predicting World States, Rendering Geometry, Painting Appearance Aug 19, 2026 164 Game world models that generate video directly often lose track of physical consistency over long sequences --- characters clip through walls or drift apart. Marionette fixes this by separating concerns: a neural model predicts explicit 3D skeletal states, a zero-parameter renderer handles geometry exactly, and video diffusion paints the final appearance. This structure lets developers im

Decoding the Past: An Uncertainty-Aware Deep Learning Framework for Sex Attribution in Prehistoric Hand Stencils Aug 19, 2026 162 Archaeologists have long puzzled over who made Upper Paleolithic hand stencils, since no ground truth exists and prehistoric hands differ from modern ones. This paper tackles that uncertainty head-on with a deep learning pipeline that generates multiple silhouette interpretations per stencil, then runs them through two neural network ensembles trained on thousands of contemporary hand ima

FinRank: An Evidence-Grounded Benchmark for Financial Question Answering and Retrieval over SEC Filings Aug 10, 2026 139 Financial question answering over SEC filings faces a subtle challenge: answers can be numerically correct yet grounded in wrong evidence, since similar facts recur across filing sections, time periods, and companies. FinRank introduces a benchmark of 1,185 expert-authored questions with gold evidence and curated hard negatives to specifically test provenance-sensitive retrieval. This is

GeoDistill-Refine: Silhouette-First Geometry Distillation for Annotation-Free Spacecraft Segmentation Aug 10, 2026 169 Segmenting spacecraft in imagery typically requires manual annotation, but foundation segmentation models can generate pseudo-masks automatically, despite geometric inaccuracies that worsen during distillation. GeoDistill-Refine improves this by stabilizing teacher predictions through prompt fusion and refining a lightweight student network using silhouette, boundary, and shape-based obje

GeoBenchLLM: A Comprehensive Benchmark for Evaluating LLMs on Geo-Related Tasks Aug 10, 2026 158 LLMs have typically been evaluated on geo-related tasks in narrow, homogeneous settings, obscuring how well they generalize across diverse geospatial and temporal challenges. GeoBenchLLM addresses this by combining twelve public datasets into a comprehensive benchmark covering varied geo-related tasks and domains. This is useful for researchers and developers building geospatial AI applic

I Seek You in Videos: Identity-Conditioned Queries for Person-Centric Video Reasoning Aug 10, 2026 148 Real-world video understanding often requires identifying and tracking specific individuals across multimodal content, a capability underserved by existing video-text benchmarks. This paper introduces the Identity-conditioned Queries task and the ISYV framework, including a challenging benchmark, large training dataset, and model designed to jointly reason over a reference image and video

ResidencyRL: Reinforcement Learning in Simulated Clinical Environments Aug 10, 2026 136 Training AI agents for complex, multi-turn clinical reasoning—like a medical resident gaining experience—remains underdeveloped despite LLMs' strong performance on static medical exam benchmarks. ResidencyRL trains clinical AI agents through simulated adversarial patient encounters spanning many dialogue turns and tool calls, rewarding diagnostic accuracy, safety, and communication qualit

CoBa: Cost-Effective Test-Time Scaling via Compute-Balanced Routing Aug 10, 2026 119 Test-time scaling strategies for LLM reasoning—generating more samples, longer chains of thought, or stronger verification—compete for a fixed compute budget, raising the question of where to best allocate resources. CoBa formulates this as a routing problem, first applying cheap verification broadly before directing only uncertain or high-value candidates to stronger, costlier verificati

A Picture is Worth a Thousand Tokens: How Vision Language Models Cut AI Energy Costs While Improving Accuracy Aug 10, 2026 141 LLM inference dominates AI's operational energy use, and numerical time-series data—like telecom network metrics—creates especially inefficient token-heavy inputs when represented as text. This paper shows that converting time-series data into 2D visual plots and processing them with vision-language models dramatically reduces token counts and energy consumption while improving accuracy.

TEPA: Revoking Stale Memories for Conflict-Robust Language Agents Aug 10, 2026 116 Language agents with long-term memory face a "memory pollution" problem: outdated facts remain retrievable even after the real-world situation changes, corrupting downstream reasoning. TEPA addresses this by treating memory validity as an explicit, revocable state, automatically invalidating stale precedents when contradicting evidence appears while preserving history for audit purposes.

Diffusion LLMs as Targets and Adversaries: Mechanistic Safety Exploits Aug 10, 2026 123 Diffusion-based LLMs use a fundamentally different generation process than standard autoregressive models, and their safety mechanisms are poorly understood. This paper reveals that safety alignment in diffusion LLMs is sparse and often inherited from autoregressive source models, making them vulnerable to transfer-based jailbreak attacks. The authors introduce SN-Guided Diffusion, an off

SABRE: Scalable and Automated Benchmarking of VLMs under Stress Aug 10, 2026 149 Vision-language models (VLMs) are advancing rapidly, but building benchmarks that meaningfully stress-test their weaknesses is costly and labor-intensive. SABRE offers an automated pipeline converting task specifications into structured images and question-answer pairs, using automated filtering plus human review to ensure benchmark quality and difficulty. Its SABRE-Prior instantiation sp

Post-Grokking Collapse at the Representation-Readout Interface in Muon-Trained Transformers Aug 10, 2026 150 This paper investigates why transformers trained with the Muon optimizer can "grok" (achieve sudden generalization on) modular arithmetic tasks faster than AdamW, yet later lose that generalization. Through detailed analysis of embedding/readout versus hidden-layer dynamics, the authors identify a representation-readout interface failure as the cause, distinguishing genuine circuit failur

Fisher-R1: Training LLM Agents for Reliable Hypothesis Testing Aug 10, 2026 122 LLM agents are increasingly used to automate scientific hypothesis testing, but they often make subtle statistical errors leading to invalid conclusions, even when their code execution is correct—a failure mode not captured by existing benchmarks. This paper introduces P-Bench, a benchmark of 425 hypothesis-testing tasks spanning economics, biology, and medicine, and Fisher-R1, an LLM age

PsychoAgent: An Affect-Sensitive Cognitive Architecture for Conflict-Aware Memory in LLM Agents Aug 10, 2026 112 Human memory retrieval is shaped not just by topical relevance but by emotional significance and unresolved conflict—a nuance missing from most LLM agent memory systems. PsychoAgent introduces a cognitive architecture that separates factual and affective memory, using a conflict-aware controller to surface emotionally salient information alongside topically relevant content. This could im

Blast Radius Aug 10, 2026 124 Agentic coding systems face rising costs from wasted context and tokens as sessions grow long. Blast Radius introduces a predictive memory management system that estimates how far an incoming prompt will reach across context and code, enabling reversible eviction of unneeded information while identifying recurring redundant transcripts for permanent removal. This has direct applications f

Taxonomy-Driven Analysis of Open-Source AI Risk Mitigation Tools Aug 10, 2026 157 As enterprises deploy LLMs at scale, governance and risk management become harder to handle manually, especially given a fragmented landscape of evaluation, security, and monitoring tools. This paper maps 21 open-source AI risk tools to 32 subcategories of the MIT AI Risk Mitigation Taxonomy using an LLM-assisted RAG pipeline analyzing code and documentation. This work is valuable for ent

SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient Descent Aug 10, 2026 155 LLM agents that handle recurring tasks often build up reusable "skills"—textual knowledge stored without retraining weights—but current methods for refining these skills lack proper diagnostic feedback and treat deletion carelessly. SkillProx introduces a proximal-gradient-inspired process that diagnoses failures, rolls back unsuccessful edits, and selectively consolidates or removes skil

Strategy-first synthesis planning for complex natural products Aug 10, 2026 162 Designing synthesis routes for complex natural products is a highly creative, expert-level chemistry task that existing algorithmic tools struggle with, since they rely on catalogued reactions unsuited to novel, densely functionalized molecules. SynthEx, an LLM-based agentic framework, proposes and critiques multiple synthesis strategies, producing routes that expert chemists rated compar

Interaction Creates Dynamical AI Behavior Absent in Isolation Aug 10, 2026 154 As AI agents increasingly interact with each other in real-world deployments, understanding emergent behaviors from these interactions becomes critical. This paper studies what happens when one AI ("boss") issues directives to another ("subordinate") without listening to replies, finding that the subordinate enters an unexpected behavioral state distinct from both its solo behavior and it

CoinRAG: Contextualized Information Nugget KV Cache Reuse for Long-Context RAG Aug 10, 2026 130 Long-context retrieval-augmented generation systems often reuse KV caches at the chunk level for efficiency, but this approach retains noisy, redundant information within coarse chunks. CoinRAG instead identifies fine-grained, query-relevant "nuggets" within retrieved chunks and reassembles their cached representations into a compact, semantically focused context. This is useful for appli

CreativeInstruct: Scalably Teaching LLMs to Balance Quality, Creativity, and Diversity Aug 10, 2026 126 Post-training typically boosts LLM quality but sacrifices output diversity and creativity, hurting both explicit creative tasks like story writing and implicit ones like reinforcement learning exploration. CreativeInstruct addresses this by teaching models to inject special markers that bias generation toward creativity while preserving post-trained quality, eliminating the need for multi

TFGformer: Multivariate Time Series Forecasting via Time-Frequency Graph Learning and Covariate Fusion Aug 4, 2026 132 Forecasting large-scale IoT sensor data over long time horizons is critical for maintenance and scheduling, but current time-series foundation models rely on static learned patterns without accessing relevant historical examples at inference time. CrossRAG solves this with retrieval-augmented forecasting: shape-aware memory retrieval robust to magnitude differences, contrastive learning t

Self-Play Meets Skill Evolution: Self-Evolving Search Agents that Pose, Solve, and Remember Aug 4, 2026 157 Self-play lets AI agents generate their own training problems, but without persistent memory, past failures don\'t shape future practice in a lasting way. SESA introduces an evolving skill memory system where a \"challenger\" poses problems, a solver retrieves relevant skills, and failures get distilled into reusable skills written back to memory --- creating a feedback loop where task di

DreamQAS: Learning a Decision-Useful World Model for VQE-Efficient Quantum Architecture Search Aug 4, 2026 151 Reinforcement-learning-based quantum architecture search is expensive because it repeatedly runs costly quantum simulations (VQE) after every circuit change, even though circuit construction itself is fully deterministic. DreamQAS improves efficiency by only learning to predict the expensive post-simulation feedback, using an ensemble model for uncertainty-aware planning and selective rea

From Code Review to Code Critique: Intent, Drift, and Spotlight for AI-Generated Diffs at Scale Aug 4, 2026 142 AI coding agents now produce more code than human reviewers can realistically evaluate, and existing AI review tools focus on trivial style issues rather than correctness or security. ARCTIC reframes this with three capabilities: predicting developer intent from conversation logs, detecting drift between intent and generated code, and spotlighting the diff regions most needing human atten

TerraNova: A Foundation Model for the Anthropocene Aug 4, 2026 151 Modeling Earth and human society together is hard because physical data is continuous and border-agnostic, while societal data is reported by country --- a geometric mismatch existing models handle poorly. TerraNova solves this with dedicated encoders for location, country, time, and task, fusing them via cross-modal transformers and a hypernetwork-generated decoder that outputs uncertain

ARB: A Matched Authorship-Rewriting Benchmark Dataset for AI-Text Detector Evaluation Aug 4, 2026 151 AI-text detectors are typically evaluated on human-versus-LLM text, but this may not reflect real-world cases where humans revise their own writing using LLMs. ARB tests this directly, creating matched text variants --- human-written, LLM-generated, human-text rewritten by LLMs, and LLM-text rewritten by LLMs --- across multiple source datasets and generator models. Results show detectors

AMTFV: Agentic Mathematical Tool-Flow Verification for LLM Self-Correction Aug 4, 2026 130 LLMs are good at solving math problems but poor at reliably verifying their own answers, since natural-language self-reflection lacks precision and code-generation approaches couple reasoning too tightly to implementation. AMTFV resolves this with a Mathematical Tool Flow interface that separates verification logic from execution: a verification agent builds a workflow and sends structure

COntExt: Towards Context-Aware Ontology Extension from Operational Metrics Aug 4, 2026 141 Organizations track vast operational metrics, but connecting these structured metric catalogues to formal ontologies is typically manual and labor-intensive, leaving valuable domain knowledge disconnected. COntExt automates this by using metric definitions and their context to suggest how new concepts should be integrated into existing ontologies, breaking the task into parent-class predi

LEMUR: Learning to Align with Multi-Objective Reinforcement Learning from Preference Feedback Aug 4, 2026 128 Real-world decision-making often involves competing goals --- like performance versus efficiency --- where defining a single reward function is difficult or impossible. LEMUR addresses this by combining multi-objective reinforcement learning with preference-based feedback, letting an agent learn from multiple humans\' preferences rather than requiring predefined reward functions. It joint

MOT-SR: Multi-Objective Tool-Augmented Scientific Equation Discovery with Large Language Models Aug 4, 2026 151 Discovering scientific equations from data is central to modeling physical systems, but existing LLM-based symbolic regression methods often ignore variable relationships and optimize only for fitting accuracy, causing premature convergence. MOT-SR solves this with a multi-objective framework balancing accuracy, complexity, and generalization, using collaborative LLM modules that select a

DungeonBench: A Benchmark for Rules-Rich Tactical Reasoning in Dungeons & Dragons Combat Aug 4, 2026 125 Many AI benchmarks reduce complex decision-making to simplified game mechanics, missing how real tactical reasoning juggles geometry, timing, resources, and interacting rules simultaneously. DungeonBench tackles this using Dungeons & Dragons combat, covering rich official ruleset content and testing agents on both single encounters and multi-encounter \"days\" requiring resource manag

TraceViT: Grounded Trace Supervision for Visual Abstract Reasoning Aug 4, 2026 148 Visual reasoning benchmarks like the Abstraction and Reasoning Corpus test whether models can infer abstract transformations from a few examples, but training typically only checks the final answer, ignoring how a model reasons through intermediate steps. TraceViT addresses this by training looped visual reasoners on step-by-step transformation chains derived from verified programs, groun

FriendBench: Benchmarking Dyadic Familiarity Inference in Humans and Multimodal Large Language Models Aug 4, 2026 123 Understanding social relationships from brief interactions is a subtle human skill; this paper asks whether AI models can match it. FriendBench tests whether models can tell if two people in a 20-second video clip are already friends or just meeting, using text, audio, and video across 26 models from seven companies. While top models match human accuracy, they show a different bias --- le

A Human-Centered Validation of the Explainability-Performance Coefficient Aug 4, 2026 125 As deep learning enters high-stakes domains like healthcare and finance, trustworthy explanations become essential --- yet measuring whether an explanation is actually good remains unresolved. This paper validates an EPC score that balances how sparse a feature explanation is against how much model performance it preserves. Tested across tabular, text, and image data, higher EPC scores al

When Does On-Policy Interaction Help? Representational Tradeoffs in Value-Based Imitation Learning Aug 4, 2026 148 Imitation learning trains agents from expert demonstrations, powering robotics and language-model training, but standard behavior cloning suffers from compounding errors when the learner can\'t perfectly mimic the expert. This paper explains why querying the expert interactively helps: it lets the learner target the expert\'s value function rather than the harder task of replicating its e

CENDRe: Concept Extraction with Natural Domain Representations Aug 4, 2026 125 Neural networks used for time-series classification---like detecting mechanical faults from sensor data---are often black boxes, making it hard to trust their predictions in safety-critical settings. CENDRe improves interpretability by extracting \"concepts\" from a CNN\'s internal representations, automatically determining how many concepts exist and precisely localizing them in both tim

The Theoretical Foundation of Socratic Tests: Dynamic, Multimodal, Conversational Examinations Aug 4, 2026 135 Traditional exams penalize ambition through subtractive grading, while oral exams introduce anxiety and power imbalances between examiner and student. This paper proposes a theoretical framework for \"Socratic Tests\" --- AI-mediated conversational exams that adaptively probe a student\'s understanding. By combining Dynamic Assessment, Bloom\'s and SOLO taxonomies, and graduated scaffoldi

AgentHPOBench: A Benchmark For Evaluating LLM Agents as Sequential Hyperparameter Optimizers Aug 4, 2026 132 As LLMs take on roles beyond code generation, acting as autonomous scientific agents, we need ways to test whether they can actually run experiments intelligently. AgentHPOBench evaluates this directly by having agents sequentially tune machine learning hyperparameters, observing results and logs before deciding on the next configuration. Testing 12 agents across 30 tasks, the benchmark e

Development of FDD-ON: an Ontology for VAV HVAC System Fault Detection and Diagnostics Aug 4, 2026 137 Buildings lose efficiency and reliability when HVAC faults go undetected, but fault-detection systems are hampered by inconsistent, siloed data across equipment and vendors. FDD-ON addresses this by creating a formal ontology for variable air volume HVAC systems, defining faults, symptoms, and impacts through a shared vocabulary and explicit cause-effect relationships. This gives machines

ExtractBench: A Benchmark for Schema-Guided Enterprise Document Extraction Aug 4, 2026 134 Enterprises increasingly deploy AI agents to pull structured data from documents like invoices, contracts, and forms --- but until now, no benchmark scored accuracy, completeness, grounding, and cost together. ExtractBench fills this gap with nearly 5,000 pages across 370 documents spanning 8 business domains and 67 document types. It reveals a key tradeoff: commercial vision-language mod

HiKV: Hierarchical Importance-Aware KV Cache with Hardware Acceleration for LLM Decoding Jul 31, 2026 123 Long-context LLM inference is bottlenecked by ever-growing KV cache memory during decoding. HiKV introduces an algorithm-hardware co-design that compresses the cache hierarchically: first evicting unimportant tokens within a budget, then loading only significant elements of retained tokens, achieving compression beyond single-granularity methods. A dedicated accelerator with a reconfigura

SceneActBench: Can Agents Act on the 3D Scenes They See? Jul 31, 2026 146 Vision-language model agents increasingly need to act on 3D scenes, not just describe them, yet existing benchmarks mostly score text responses or single-object tasks. SceneActBench introduces a unified agent-environment loop evaluating five 3D action tasks across 210 source instances and 520 task cases, using images, video frames, and 3D assets, scored against hidden ground truth with ge

A Self-Calibrating Agentic AI Framework for Autonomous Edge Resource Allocation Jul 31, 2026 126 As LLM agents move into autonomous, tool-executing roles, reliability suffers from lack of ground truth and operational drift in open-ended environments. This framework introduces a self-calibration mechanism using an ARIMA forecaster to approximate ground truth without continuous human oversight, applied to profiling zero-knowledge workload resource usage in edge computing networks. It i

PRIMS: Physics-guided Representation for Fluid Identification in Multimodal Sensing Jul 31, 2026 154 On-device fluid identification for microfluidic applications is challenging under varying flow, pressure, and temperature, and existing learning methods ignore underlying physics. PRIMS addresses this with a physics-aware multimodal Transformer combining physics-based sensor token vectorization, a viscosity-aware component synthesizer, and physics-guided attention fusion, embedding fluid

Unboxing Diffusion Models for the Arts: Interactive Model Bending and Practice-Based Explainability Jul 31, 2026 142 Large text-to-image diffusion models are usually opaque black boxes, limiting artists' ability to creatively manipulate them. This work reframes explainable AI for creative practice, proposing hands-on "model bending" through an interactive inspection interface built into ComfyUI, enabling layer selection and intervention. Applied to Stable Diffusion 1.5, the authors show that manipulatin

Robot Learning to Communicate through Projected Visual Abstractions Jul 31, 2026 116 Robots typically communicate only through physical movement, unlike humans who also use shadows and silhouettes. This work builds a robotic system with a 21-degree-of-freedom soft-skinned hand and a learned shadow self-model that maps hand configurations to projected shadow appearance. Given a target shadow image or video, the robot optimizes hand poses through gradient-based search and c

Hyperball May Not Be a Free Lunch Jul 31, 2026 123 Hyperball-style optimizers, which normalize updates for scale-invariant networks, have shown strong large-scale training performance, but why remains unclear. This work derives an angular effective learning rate accounting for update angle, parameter norm, and update norm, then decomposes updates into radial and tangential components to explain optimizer behavior differences like why Muon

Dynamic Capability Scoping for Enterprise AI Agents: A Synthetic Dataset and Three-Source Permission Architecture Jul 31, 2026 161 Enterprise AI agents often hold static, overly broad credentials for every possible task, expanding security risk. This work proposes dynamic least-privilege scoping via a three-source architecture combining role-based ceilings, a task-context classifier, and policy-derived prohibitions, ensuring agents only access credentials relevant to their current task. The authors release a validate

Phylogenetic signal in marine mammal and bird vocalizations captured by audio foundation models: the limited benefit of domain-specific pretraining Jul 31, 2026 164 This study asks whether general-purpose audio embeddings implicitly encode evolutionary relationships never explicitly trained into them. Testing four pretrained audio models on marine mammal and bird vocalizations, the authors found strong correlations between embedding-space distances and phylogenetic distance, particularly among cetaceans, robust even after controlling for frequency an

Beyond Perspectives: A Trio-Ethnography of Interpretation Evolution in LLM-Supported Programming Education Jul 31, 2026 164 As generative AI reshapes programming education, instructors often infer student learning purely from classroom observation, missing crucial context. This study uses trio-ethnography, pairing two educators with differing teaching philosophies alongside an undergraduate student, across three dialogic sessions exploring AI-assisted learning. The student's firsthand narratives revealed learn

TRACE-ROUTER: Task-Consistent and Adaptive Online Routing for Agentic AI Jul 31, 2026 158 Enterprise AI deployments often route each LLM call independently to balance cost and quality, but agentic workflows only get evaluated by delayed, task-level outcomes, misaligning per-call routing with actual performance signals. TRACE-Router fixes this by assigning an entire task to one model at the start via a contextual bandit, then updating its policy using the task's terminal reward

Learning to Prepare Molecular Ground States with Transformer Models Jul 31, 2026 141 Quantum chemistry simulations require efficient state-preparation circuits, but iterative methods like ADAPT-VQE become computationally prohibitive for large, pharmaceutically relevant molecules. ADAPT-GQE addresses this with a generative AI framework trained on ADAPT-VQE-generated reference circuits, then improved further via reinforcement learning to exceed the training data's accuracy.

MineValiCoder: Reliable Code Generation with Test Case Quality Mining and Bipartite Graph-Based Mutual Validation Jul 31, 2026 158 LLM-based test-driven code generation struggles when only natural-language requirements exist, since automatically generated tests can be faulty or inconsistent, misleading optimization. MineValiCoder tackles this with a closed-loop framework: filtering unreliable tests via self-validation, iteratively refining diverse code candidates, and using a bipartite graph model to jointly validate

k-LoRA: Condition Numbers Reveal Which LoRA Matrices Worth Updating Jul 31, 2026 151 Low-Rank Adaptation (LoRA) fine-tunes large models efficiently but conventionally updates all matrices uniformly, wasting compute on matrices that contribute little. κ-LoRA shows that matrices with higher condition numbers hold underdeveloped directions driving most adaptation gains, while low-condition-number matrices are already balanced and add little value. By restricting updates to t

CausalForge: A Formally Grounded, Self-Improving Agentic Framework for Automated Research in Causal Inference Jul 31, 2026 153 LLM-driven automated research is limited by unreliable evaluation, since LLM reviewers can accept fabricated results near chance levels. CausalForge addresses this by grounding causal inference research in the Lean proof assistant, pairing a machine-checked causal inference library (Causalean) with an agentic pipeline (CausalSmith) that proposes, formalizes, and proves results, then audit

Opaque Epistemic Mediation: How LLM Deployment Configurations Shape the Validation of Pseudo-Science Jul 31, 2026 167 This study probes how commercial LLMs (Claude, Grok, GPT, Gemini) evaluate ethnonationalist pseudo-scientific claims across time and interfaces, finding Grok's default versions rated such claims far more credible than competitors, with behavior shifting via undocumented silent patches and diverging between API and web access for the same model. The findings reveal that an LLM's epistemic

Quantum Spectral Model: Data Reuploading with Input-Conditioned Frequency Support Jul 31, 2026 128 Quantum machine learning models typically encode matrix-valued data using generic rotation gates that ignore matrix-level spectral structure. Quantum Spectral Models (QSMs) instead build the data-encoding unitary's generator directly from each input's spectral properties, testing symmetric, global-block, and patch-local Hamiltonian variants. Evaluated on Pendigits and synthetic spectral-s

The Regression Tax: Decomposing Why Skills Help and Hurt LLM Agents Jul 31, 2026 159 Adding "skills" (procedural guidance) to LLM agents is usually judged by average performance gains, but this masks cases where skills actively cause failures on previously solvable tasks. Analyzing nearly 6,000 runs across office automation benchmarks, the authors identify three regression mechanisms: skill presence alone altering behavior, skills overriding correct input interpretation,

Explainable Reinforcement Learning for assisting Air Traffic Controllers Jul 31, 2026 147 As AI moves into safety-critical domains like aviation, healthcare, and autonomous driving, trust hinges on explainability. This work applies explainability techniques to a reinforcement learning agent trained in a simplified air traffic control environment, where the agent chooses alternative flight routes to avoid no-fly zones. Using saliency maps, the authors expose which input feature

SM4RT: Learning Structured Motion Geometry for 4D Reconstruction Jul 31, 2026 162 Most monocular motion-perception systems treat every pixel's movement as an independent displacement, ignoring that real objects move as rigid bodies. SM4RT challenges this by encoding scene dynamics as a small set of shared "motion bases," each a temporal sequence of 6D twists in SE(3), so points belonging to the same object inherit a common rigid trajectory. A parallel encoder-decoder r

Episodes

Recommended