Briefing

DR-Gym is an open-source Gymnasium-compatible environment that trains and evaluates demand-response from the electric utility perspective, addressing the missing feedback loop in offline smart-meter datasets. The work pairs physics-based building demand profiles with a regime-switching wholesale price model calibrated to real-world extreme events and a configurable multi-objective reward. Unlike device-level simulators, DR-Gym targets market-level utility decision-making and demonstrates realistic, learnable scenarios using baseline strategies.

Rethinking energy resilience for AI infrastructure (DCD Rehlko session, published 11 May 2026) argues power is now the limiting factor for data center growth. It advocates integrated backup and continuous‑power architectures, fuel flexibility across natural gas, HVO, hydrogen, RNG and BESS, and explicit trade‑offs between uptime, efficiency, curtailment and emissions in site design and operations.

The eBook 'Redefining the data center for AI' (DCD Edge Infra & Inference Channel, published 13 May 2026 on datacenterdynamics.com) compiles insights from Schneider Electric and partners, citing an Innovation Summit estimate that 88% of companies pursue AI, and provides practical guidance on designing AI-ready data centres—power, cooling, density, efficiency and scalable architectures.

Based on the abstract (full text not consulted), the paper analyzes how routing decisions in Sparse Mixture-of-Experts form a geometric coupling between routers and experts: gradients for a routed token point along the same input direction in router and expert weights. The authors validate this in a 1B-parameter SMoE, show that auxiliary load-balancing disrupts the coupling (making router directions ≈3× more similar), and introduce a parameter-free online K-Means router (running-average centroids + cosine assignment) that minimizes load imbalance with only modest perplexity cost, suggesting geometric coupling underlies effective specialization. Published 2026-05-12 by Ahrac, Hochwald, and Geva (arXiv:2605.12476v1).

Attractor Models introduce a two-stage iterative-refinement architecture where a backbone proposes embeddings and an attractor module solves for a fixed point with gradients via implicit differentiation, which keeps training memory constant and allows adaptive iteration depth. Empirically (abstract only), they report up to 46.6% perplexity reduction and 19.7% downstream accuracy gains in pretraining—e.g., a 770M model beating a 1.3B Transformer trained on twice the tokens—and strong few-shot reasoning (27M params, ~1,000 examples: 91.4% Sudoku-Extreme, 93.1% Maze-Hard). The paper also describes 'equilibrium internalization,' where fixed-point training moves initial outputs near equilibrium so the solver can be removed at inference with little loss. Full text was not available here (abstract only).

KV-Fold introduces a training-free long-context inference protocol that treats the transformer's KV cache as an accumulator in a left fold, appending newly produced keys/values per chunk and forwarding the enlarged cache. The method yields a stable, precision-robust recurrence (plateau insensitive to 10,000× precision changes) and achieves 100% exact-match retrieval on Llama-3.1-8B across 152 trials (16K–128K tokens, depth ≤511). Based on the abstract (full text not provided).

ORCE introduces a decoupled, order-aware approach to verbalized confidence: answers are generated first and confidence is estimated conditioned on the fixed question–answer pair. Using a sampling-based surrogate of multiple completions and rank-based RL to encourage higher confidence for more likely-correct responses, the method improves calibration and failure prediction without substantially hurting answer accuracy, addressing interference from prior joint optimization techniques.

Multi-Stream LLMs (Su et al., 2026) argue that single-stream chat/instruction formats (e.g., instruction-tuned ChatGPT-style models) create a bottleneck preventing simultaneous reading, thinking and acting. They propose instruction-tuning for parallel computation streams—splitting roles into separate input/output channels so each forward pass concurrently consumes and emits tokens across streams. The preprint (37 pages) reports this architecture should enable concurrent action, boost efficiency via parallelization, and improve security and monitorability; code is provided on GitHub.

The paper proposes Self-Supervised Laplace Approximation (SSLA), which sidesteps parameter posteriors and directly approximates the posterior predictive by refitting on model self-predictions; a faster approximate variant (ASSLA) avoids costly refits. The authors provide theoretical analysis and experiments on Bayesian linear models through Bayesian neural networks, showing better predictive calibration than classical Laplace methods while keeping computation efficient.

VECA (Visual Elastic Core Attention) tackles the quadratic cost of ViT self-attention by introducing a small set of learned core embeddings (C) that mediate communication for N patch tokens, producing linear O(N) complexity for fixed C. Cores are initialized and propagated across layers; nested core-axis training permits elastic compute–accuracy trade-offs. The authors claim competitive results on classification and dense tasks with lower compute (see arXiv and GitHub).

Reward hacking in rubric-based reinforcement learning: Mahmoud et al. (arXiv 2026-05-12) analyze divergence between training verifiers and a three-judge reference panel across medical and science domains. They separate verifier failure from rubric-design limitations, show weak verifiers yield nontransferable gains and recurring exploitation, and propose the self-internalization gap (policy log-probabilities). Stronger verification reduces but does not guarantee broader quality gains. (Summary based on abstract only.)

Pion introduces an orthogonal equivalence transformation optimizer that modulates weight-matrix geometry by applying left/right orthogonal updates while strictly preserving singular values and the spectral norm. The paper derives the Pion update rule, studies design options and convergence properties, and presents empirical evidence that Pion is a stable, competitive alternative to additive optimizers (e.g., Adam, Muon) for LLM pretraining and fine-tuning.

SenseNova-U1 introduces NEO-unify, a unified multimodal paradigm that treats understanding and generation as a single process. The authors release two models (8B dense and 30B MoE A3B) and claim parity with top-tier understanding-only VLMs across perception, reasoning, decision-making, and spatial tasks, while also achieving strong any-to-image synthesis and interleaved multimodal generation; full design and training details are provided on the project page.

LongMemEval-V2 (LME-V2) is a benchmark for assessing whether memory systems let agents acquire environment-specific experience; it provides 451 questions across five memory abilities with histories up to 500 trajectories (115M tokens). The authors evaluate a RAG-style AgentRunbook-R and a file+coding-agent AgentRunbook-C, reporting 72.5% accuracy for AgentRunbook-C versus 48.5% for the best RAG baseline and 69.3% for a coding-agent baseline, while noting higher latency; summary based on the abstract.

The Utility Dive Studio report (published 2026-05-13) synthesizes views from 135 energy executives on balancing growth, cost, and sustainability amid accelerating electrification and looming decarbonization targets. Respondents describe managing volatility, redirecting investment priorities, and maintaining strong confidence in the energy transition; the analysis is custom sponsored content by Utility Dive’s Studio.

Anthony Sagami (@anthonysagami) shared a WSJ article on 2026-05-13 describing orbital data centers: swarms of AI-chip satellites with solar arrays in near-polar orbits to maximize sunlight. He appended a direct investment pitch, listing 49 AI-related tickers (e.g., $NVDA, $MSFT, $META, $GOOGL, $AMZN) and urging buys.

SI-Diff addresses contact‑rich assembly (peg‑in‑hole) by learning a single force‑domain diffusion policy with a mode‑conditioning mechanism and a search teacher policy that generates diverse trajectories; it maps tactile and end‑effector velocity observations to actions. Experiments report extending x–y misalignment tolerance from 2 mm to 5 mm versus TacDiffusion and strong zero‑shot transfer to unseen shapes. Abstract only.

DexTwist introduces a functional twist-retargeting method for MR-based teleoperation that targets contact-rich rotational manipulation where kinematic imitation fails. The system detects tripod pinches, estimates intended screw axis and twist, then performs a real-time joint-space residual optimization minimizing a virtual-object objective (turn angle, axis consistency, fingertip closure, tripod stability). Simulations and real tests demonstrate improved turning-angle tracking and reduced screw-axis drift versus a vector-based baseline.

CUActSpot targets the long-tail of complex GUI interactions that undermine computer-use agents by providing a multimodal benchmark (GUI, text, table, canvas, natural image) and a renderer-based data-synthesis pipeline: automatic scene generation, screenshot/element-coordinate recording, and LLM-produced instructions/action traces. Training on this corpus yields Phi-Ground-Any-4B, which outperforms open models under 32B parameters. Only the abstract was available for this summary.

A popular dining-focused credit card received an anniversary refresh on May 11, 2026 that adds a 5X hotel-earning rate, nearly $100 in limited-time travel and dining credits, and rental-car elite status (enrollable now). AwardWallet notes the upgrade is coupled with a welcome bonus up to 100,000 points, though the email withheld the issuer/card name for advertising reasons.

Task-Adaptive Embedding Refinement via Test-time LLM Guidance presents a test-time method that refines query embeddings via a generative LLM's feedback on a small document subset to tailor embeddings to ad-hoc zero-shot search and classification tasks. Experiments with state-of-the-art embedding models across diverse benchmarks report consistent gains (up to +25% relative), improving ranking and class separation; code released on GitHub. Abstract only; full text not provided here.

The paper introduces a proximal-gradient Monte Carlo sampler for composite log-concave densities π ∝ e^{-f-g}, combining gradient steps on smooth f with a restricted Gaussian-oracle (RGO) proximal sampler for g. Under α-strong convexity of f+g and β-smoothness of f it achieves ε total-variation error in ~O(κ·sqrt(d)·log^4(1/ε)) iterations (κ=β/α), and the authors extend guarantees to Poincaré/LSI targets and to Lipschitz non-smooth f.

Ramp's merch fulfillment service (announced in an email published 2026-05-11) combines printing with inventory storage, pick-and-pack operations and worldwide shipping, plus online store setup, onboarding kits and event/campaign distribution. The offering is modular — clients can select full fulfillment or specific services — and was credited with selling out The Bugle Podcast's limited-edition Christmas jumpers after Ramp printed and shipped orders globally.

MoLA (Mixture of Latent Actions) targets the gap between video-based imagination and actionable control: instead of feeding predicted frames to a policy or decoding videos directly into controls, it infers a mixture of latent actions via pretrained, modality-aware inverse-dynamics models (semantic, depth, flow) to produce a physically grounded action interface. Evaluated on LIBERO, CALVIN, LIBERO-Plus and real robots, the abstract reports consistent improvements in success rates, temporal consistency, and generalization; summary based on the abstract (full paper not reviewed).

EgoEV-HandPose tackles egocentric 3D bimanual hand-pose estimation and gesture recognition from stereo event cameras by introducing KeypointBEV, which lifts stereo features into a bird's-eye-view and iteratively reprojection-refines depth and kinematic estimates. Trained and evaluated on the new EgoEVHands dataset (5,419 sequences, 38 gestures), it reports MPJPE 30.54 mm and 86.87% Top-1 accuracy, outperforming RGB-stereo and prior event-based methods, notably under low-light and occlusion.

Laplacian-based neural operators are analyzed for the generalized Gierer–Meinhardt reaction–diffusion system: the paper obtains explicit approximation-error bounds depending on network depth, width, and spectral rank by using the Laplacian eigenfunction expansion of the PDE Green’s function. The authors show parameter complexity scales at most polynomially with accuracy and present numerical experiments consistent with theory. Summary based on the abstract; full text not reviewed.

The paper tackles low-latency, whole-body humanoid teleoperation by mapping Virdyn IMU suit data to a Unitree G1 using a custom motion-processing, kinematic-retargeting, and control pipeline that avoids offline buffering and learning-based modules. Validated in MuJoCo then transferred unchanged to the physical robot, the system reportedly achieves stable, synchronized reproduction of a wide motion repertoire; summary based on the abstract and metadata.

TMRL and Context-Smoothed Pre-training (CSP) inject forward-diffusion noise into policy inputs during pretraining to create a continuum from precise imitation to broad action coverage, then train agents to modulate the diffusion timestep during RL fine-tuning to control exploration. The method works with states, 3D point clouds, and visual policies and enables sub-hour real-world manipulation fine-tuning; full paper and code on arXiv and project site.

The paper develops a model-based bootstrap for transition kernels in finite controlled Markov chains with possibly nonstationary or history-dependent policies, proving distributional consistency in both long-chain and episodic offline RL regimes. Using a novel bootstrap LLN and a martingale CLT, the authors extend results to OPE and OPR via Hadamard-differentiable Bellman operators, producing asymptotically valid CIs; RiverSwim experiments show strong empirical calibration.

The paper introduces the first online L2D algorithm for multiclass classification with bandit feedback and a dynamically varying expert pool, addressing streaming data and shifting expert availability. It achieves regret O((n+ne)T^{2/3}) generally and O((n+ne)√T) under low noise, relying on new H-consistency bounds and first-order online convex optimization; experiments validate practicality. Summary based on the abstract (full text not available).

Siebenborn et al. formalize bilateral morphological symmetry in bimanual mobile manipulation and propose a C2-equivariant flow-matching policy that enforces reflection symmetry through loss regularization or an equivariant velocity network. On planar and 6-DoF tasks the method boosts sample efficiency and enables zero-shot generalization to mirrored states, with real-world TIAGo++ validation. Summary based on the abstract; full text not reviewed.

Optimal policy learning under combined budget and minimum-coverage constraints is treated as a knapsack-type allocation problem; Cerulli (May 2026) proves the optimal rule is an affine threshold in budget and coverage shadow prices, shows an LP relaxation has an O(1) integrality gap, and evaluates two algorithms (GLC, RC), with Monte Carlo confirming near-optimal finite-sample performance and predictable failure modes.

TriBand-BEV introduces a fast LiDAR-only 3D pedestrian detector that encodes the full point cloud into a lightweight 2D BEV tensor with three height bands, reformulating 3D detection as 2D detection and reconstructing boxes post-hoc. Using area attention, a hierarchical bidirectional neck (P1–P4), and a distribution-focal rotated-IoU head, it reaches 58.7/52.6/47.2 BEV AP on KITTI at 49 FPS; code is public.

TextSeal presents a practical, localized watermark for LLM outputs that combines Gumbel-max sampling, dual-key generation, entropy-weighted scoring, and multi-region localization to preserve diversity while enabling strong provenance detection. The method adds no inference cost, supports serving optimizations like speculative decoding, strictly dominates prior baselines (e.g., SynthID-text), is robust to dilution, transfers through distillation, and a 6,000 A/B multilingual study (5 languages) reported no perceptible quality change. Full paper text was not available in the provided content.

AmbiSuR revisits Gaussian Splatting to improve photometric‑ambiguity‑robust 3D surface reconstruction. The authors uncover two primitive‑wise ambiguities and an intrinsic self‑indication ability in the representation, then introduce photometric disambiguation and an ambiguity‑indication module to constrain and correct geometry. Experiments reportedly yield superior reconstructions across challenging scenes; paper on arXiv (2026-05-12) and accepted at ICML 2026.

Multi-Variable Conformal Prediction (MCP) tackles the limitation of conventional conformal methods that use a scalar score and single threshold by allowing vector-valued scores and multiple calibration variables. Using scenario theory, MCP unifies prediction-set design and calibration into one optimization problem (no data split) and provides finite-sample coverage. Two variants, RemMCP and RelMCP, trade off convexity assumptions and conservatism; experiments on ellipsoidal and multi-modal sets show target coverage, smaller/comparable set sizes, and lower calibration variance. Full text on arXiv.

FuTCR (Future-Targeted Contrastive and Repulsive) tackles Continual Panoptic Segmentation by preventing the collapse of diverse unlabeled objects into a single background representation. The method groups predicted masks with background labels but non-background logits to find future-like regions, uses pixel-to-region contrast to form coherent prototypes, and repels background features from known-class prototypes. According to the abstract, FuTCR yields up to 28% relative gains on new-class panoptic quality while maintaining or improving base-class performance (up to 4%), evaluated across six CPS settings and varied dataset sizes.

MEME (Multi-entity & Evolving Memory Evaluation) targets LLM-agent failures when storing, updating, and reasoning about many entities across sessions. The benchmark defines six tasks (including Cascade, Absence, Deletion) and tests six memory systems across three paradigms on 100 controlled episodes. Results show catastrophic collapse on dependency reasoning (Cascade 3%, Absence 1%), and only an expensive file-based agent + Claude Opus 4.7 partially closes the gap, highlighting a practical-performance tradeoff.

EgoForce tackles depth–scale ambiguity and device-specific generalization in monocular, head-mounted hand capture by fusing a differentiable forearm model, an arm–hand transformer that predicts geometry from a single egocentric view, and a ray-space closed-form solver to recover absolute camera-space 3D pose. The method works across fisheye, perspective, and wide-FOV optics and yields up to 28% MPJPE reduction on HOT3D, with code and data released.

The Algorithmic Caricature (Gunjan et al., arXiv 2026-05-12) evaluates whether LLM-generated political posts replicate real online populations by comparing a paired corpus of 1,789,406 posts across nine crisis events. It finds synthetic text is fluent but population-level unrealistic—more negative, less sentiment-dispersed, structurally regular, and lexically abstract—with gaps varying by event and summarized by a proposed 'Caricature Gap'. Full text not available; summary based on abstract.

OmniNFT (Zhang et al., arXiv 2026-05-12) targets RL fine-tuning for joint audio–video generation by diagnosing three failure modes—advantages inconsistency, gradient imbalance, and uniform credit assignment—and proposing modality-wise advantage routing, layer-wise gradient surgery, and region-wise loss reweighting in an online diffusion-RL pipeline. Tests on JavisBench and VBench with the LTX‑2 backbone report improved per-modality quality, cross-modal alignment, and synchronization. Summary based on the abstract.

The paper presents a real-world dataset aimed at enabling AI-native mobility in 6G by replacing common simulation-based data with measurements from a commercial network. To address high interruption times and measurement overhead during UE mobility, the authors collected multi-speed traces across pedestrian, bike, car, bus, and train scenarios, emphasizing handover events. A key contribution is inclusion of timing advance (TA) at RACH trigger, MAC CE, and PDCCH grant events. The authors provide dataset generation details and exploratory analyses and propose use cases such as TA prediction and AI/ML-driven beam and handover management (arXiv:2605.12453v1).

Perception Deep Research frames open-world visual perception where target identities must be resolved from external web facts before localization. The authors introduce WebEye — a benchmark with 120 images, 473 annotated objects, 645 QA pairs and 1,927 task samples — and propose Pixel-Searcher, an agentic search-to-pixel workflow that attains top open-source results across grounding, segmentation, and VQA.

Shen and Kuusela (2026) introduce a score-augmented loss for neural likelihood surrogates in simulation-based inference, augmenting binary cross-entropy with exact parameter-space score ∇_θ log p(x | θ) and adaptive weighting based on loss gradients. Evaluated on network dynamics and spatial processes, the approach boosts surrogate quality and can match the effect of 10× more training data with under a 10% training-time increase.

GuidedVLA proposes guiding Vision-Language-Action models by supervising individual attention heads with manually defined auxiliary signals, rather than relying on end-to-end implicit learning. The paper implements three specialized heads (object grounding, spatial geometry, temporal skill logic) and reports higher success rates on simulated and real-robot tasks versus strong VLA baselines. Full text was not available in the provided abstract.