Briefing

GraphDPO, introduced by Ning Liu et al. (arXiv 2026-05-08), extends DPO to operate over directed acyclic preference graphs built from multiple rollouts per prompt. It optimizes a Plackett–Luce–inspired, graph-aggregated objective, uses equivalence-class layers to avoid spurious gradients, keeps linear per-prompt cost via log-sum-exp, and—with optional annealed ground-truth anchoring—outperforms pairwise DPO on reasoning and program synthesis tasks.

Fast Byte Latent Transformer (BLT) variants tackle slow byte-by-byte autoregressive generation by introducing BLT-D — a block-wise diffusion-trained model enabling multi-byte parallel decoding — plus BLT-S and BLT-DV speculative/verification extensions; experiments (abstract-only) report an estimated >50% memory-bandwidth reduction, full text not available.

GLiGuard is a 0.3B-parameter, schema-conditioned bidirectional encoder (adapted from GLiNER2) that encodes task definitions and label semantics into structured token schemas to evaluate prompt safety, response safety, refusal detection, 14 fine-grained harm categories, and 11 jailbreak strategies in one non‑autoregressive forward pass. On nine safety benchmarks it matches F1 of 7B–27B decoder guards while being 23–90× smaller and yielding up to 16× throughput and 17× lower latency. Summary based on the paper's abstract only.

AutoTTS introduces an environment-driven approach that replaces hand-designed TTS heuristics with automatic discovery: width–depth TTS is cast as controller synthesis over stored reasoning trajectories and probe signals. The method uses beta parameterization and fine-grained execution-trace feedback to make search tractable and efficient, yielding better accuracy–cost tradeoffs on math reasoning tasks; discovery cost was $39.9 and 160 minutes.

Normalizing Trajectory Models (NTM) address the failure of Gaussian denoising assumptions when compressing diffusion to few steps by modeling each reverse step as a conditional normalizing flow trained with exact trajectory likelihood. The design stacks shallow invertible blocks per step with a deep parallel trajectory predictor, is trainable from scratch or from flow-matching initials, and—per the abstract—enables self-distillation that yields high-quality text-to-image samples in four steps while retaining exact likelihood.

Asymptotically Log-Optimal Bayes-Assisted Confidence Sequences for Bounded Means introduces a Bayes-assisted method that uses a Bayesian working predictive model to choose one-step martingale updates maximizing predictive expected log-growth, yielding time-uniform confidence sequences for bounded IID means. The authors prove Wasserstein-consistency of the predictive implies asymptotic log-optimality (matching an oracle per-sample log-growth). They instantiate methods with Dirichlet-process mixtures and Bayesian exponentially tilted empirical likelihood and demonstrate reduced interval widths and sampling in synthetic and real tasks (LLM best-arm, prediction-powered inference).

Data Centre Dynamics' Telecoms & Connectivity episode (published 11 May 2026) brings Belden experts to examine how industrial and edge environments must evolve connectivity to handle AI, automation and distributed infrastructure. The briefing focuses on simplifying network architectures, boosting reliability and visibility, and meeting growing data volumes across complex operational sites; viewers can register to stream the session.

MiniCPM-V 4.6 (1.3B) is a high-resolution multimodal model from OpenBMB released 2026-05-11 that applies LLaVA-UHD v4 to cut vision encoding costs by 55%, claiming superior benchmark results versus Gemma4-E2B-it and Qwen3.5-0.8B while using just 2.5% of Qwen's token budget; OpenBMB reports 75.7 ms TTFT (2.2× faster on 3136² images) and ~1.5× token throughput on an RTX 4090, with model and demos available on Hugging Face, GitHub and Modelscope.

Conformal Path Reasoning (CPR) is a trustworthy Knowledge Graph Question Answering framework that performs conformal calibration at the query level over path-level scores and adds a lightweight Residual Conformal Value Network (RCVNet) trained with PUCT-guided exploration to produce discriminative nonconformity scores. According to the abstract, CPR substantially improves empirical coverage (+34%) while shrinking prediction sets (−40%) versus conformal baselines, yielding more reliable and compact answer sets. Full text was not used beyond the provided abstract.

HFRU is a reinforcement unlearning method for VLMs that removes sensitive visual knowledge by modifying the vision encoder through a two-stage process: alignment disruption followed by GRPO optimization with a composite reward (including an abstraction reward to avoid object-hallucination). On object-recognition and face-identity tests reported in the abstract, it achieves >98% forgetting and retention with negligible hallucination. Summary based on abstract; full paper not reviewed.

TAVIS is a benchmark for egocentric active vision and anticipatory gaze in imitation learning that provides two complementary suites (Head: 5 tasks; Hands: 3 tasks) on GR1T2 and Reachy2 torsos in IsaacLab, plus ~2,200 demonstrations. It introduces a paired headcam vs fixedcam protocol, the GALT anticipatory-gaze metric, and ID/OOD splits; baselines show selective benefits from active gaze, brittleness of multi-task policies under shift, and human-like anticipatory gaze emerging from imitation.

Many-to-Many Multi-Agent Pickup and Delivery (M2M) tackles warehouse many-to-many MAPD, where SKUs can be picked up from or delivered to multiple locations, yielding an NP-hard four-dimensional assignment. The authors propose two algorithmic variants (duration-minimizing M2M and SKU-aware M2M-wSKU) and report simulations over 8-hour operations showing up to 22,000 more tasks completed on average versus prior methods. Full text was not provided with the abstract.

MiniCPM‑V 4.6 (1.3B) from OpenBMB is presented as a mobile/edge‑optimized multimodal model: LLaVA‑UHD v4 reportedly cuts vision encoding costs 55%, enabling native edge deployment. The team claims TTFT 75.7 ms (2.2× faster than Qwen3.5‑0.8B at 3136²), ~1.5× token throughput on an RTX 4090, and superior benchmark scores while using only 2.5% of Qwen’s token budget.

Bilevel minimax optimization with minimax lower-level problems is addressed via penalty-based first-order methods that remove the need for lower-level strong convexity. Based on the abstract, the deterministic method attains an ε-KKT point in tilde O(ε^{-4}) oracle calls, the constrained lower-level case improves prior tilde O(ε^{-7}) to tilde O(ε^{-4}) via Lagrangian duality, and a stochastic variant achieves near-ε-KKT with tilde O(ε^{-9}).

Varick argues AI success requires implementation services: teams that analyze and refactor processes and build custom agents, not just selling tools. He touts Varick’s record (100% production, 100% repeat clients), notes OpenAI launched an OpenAI Deployment Company on 2026-05-11 backed by 19 partners, warns of vendor lock-in risks, and is hiring Engineers, FDEs, and Consultants.

Discriminative factorization is proposed to quantify and distinguish high- versus low-quality query sets for black-box model-level classification. The framework yields a theoretical result: the probability of chance-level classification decays exponentially with query budget. On three auditing tasks the authors show estimated parameters track empirical decay and enable query selection that mirrors oracle ordering.

CUTS-GPR targets the computational bottleneck of Gaussian process regression in high-dimensional problems by exploiting an additive kernel on an incomplete grid to reveal structure that yields an extremely fast kernel matrix–vector product. The method shows near-linear (or linear) scaling in N and low-order polynomial scaling in D, with benchmarks on billions of points and thousands of dimensions and a full GPR + hyperparameter run in hours for N=447,265, D=24, enabling Bayesian modeling of high-dimensional potential energy surfaces in computational chemistry.

Distributional treatment effects that leave means unchanged are targeted by DR-ME, a semiparametrically efficient finite-location test (Zenati & Gretton, arXiv:2605.08034v1, 2026-05-08). From observational data it derives orthogonal doubly robust kernel features whose centered oracle is the canonical gradient; for fixed locations the test is chi-square calibrated and has noncentral chi-square local power with covariance whitening optimizing local SNR. Sample splitting preserves post-selection validity; experiments show near-nominal Type-I error and competitive power, with learned locations that localize effects in a semi-synthetic medical-imaging study.

The paper investigates whether a single-DOF actuated sagittal spine improves learned agile locomotion for a quadruped. Using MuJoCo simulations of the Silver Badger robot, the authors evaluate high-speed running, stair and high-angle slope climbing, hurdling, and crawling. Results show the spine increases agility and enables traversing higher stairs, steeper slopes, taller obstacles, and narrower passages, suggesting spine actuation is a promising design extension for agile robots.

Yo Ehara proposes a Householder-aligned permutation test to more accurately compare word semantic breadth from contextualized token embeddings. By reflecting one token cloud to align mean directions before permutation testing, the method prevents directional differences from masquerading as dispersion effects. Empirically it cut Type‑I error by 32.5% and, with a GPU-batched implementation, ran 23× faster than a CPU baseline; accepted to ACL 2026.

Text-to-CAD evaluation is framed as automated testing in this work: Mallis et al. propose CADTestBench and CADTests, executable checks that validate geometric and topological constraints of generated CAD models. They benchmark recent Text-to-CAD systems on CADTestBench and demonstrate that using CADTests to guide generation yields simple baselines that outperform prior methods; code and datasets are open-sourced.

Vibe-Trading is an open-source AI + quant trading framework that hit 6k+ GitHub stars and has shipped daily updates for a month (published 2026-05-11). It connects to Tushare and AKShare, supports A‑shares/crypto/global markets, exports to TradingView/通达信/MetaTrader, and offers brokerage analytics, a Shadow Account, persistent-memory agents (13+ models), and Docker-ready security for AI+Human workflows.

Proxy3D proposes compact 3D proxy representations extracted from video frames via semantic and geometric encoders and semantic-aware clustering. The authors curate the SpaceSpan dataset and apply multi-stage training to align proxies with vision-language models; on 3D visual question answering, visual grounding, and spatial intelligence benchmarks the method achieves competitive or state-of-the-art results while using shorter vision sequences. (Abstract-only summary.)

Flow-OPD addresses reward sparsity and gradient interference in multi-task alignment for Flow Matching text-to-image models by combining single-reward GRPO teachers with a Flow-based Cold-Start and on-policy distillation into a single student, plus Manifold Anchor Regularization to prevent aesthetic degradation. Built on Stable Diffusion 3.5 Medium, the abstract reports GenEval 63→92 and OCR 59→94; full text was not available.

Online kernel regression: Li and Hiratani (2026) derive a closed-form expression showing online kernel regression is equivalent to offline kernel regression with shifted, inaccurate target outputs. They give a closed-form and an iterative target-correction that provably recovers the offline predictor. Experiments on CIFAR-10 and CORe50 show online SGD with corrected targets outperforms using true targets in continual learning.

A popular dining-focused credit card received an anniversary refresh (reported May 11, 2026) that adds 5X points on hotels, nearly $100 in limited-time travel and dining credits, and enrollable rental-car elite status; the promotion also advertises a welcome bonus up to 100,000 points and requires enrollment for some benefits.

Prop trading on Polymarket is now live, according to @0xd1namit (2026-05-11), who credits @BagCalls and his team for standout marketing and expects strong execution. The Funding Predicts Beta Competition (powered by Polymarket) launches with over 1.4m in funded accounts and 10K+ in cash prizes, giving traders 14 days to qualify at fundingpredicts.com.

SCOPE proposes a specification-guided framework that tracks semantic commitments across generation by decomposing intents into structured specifications and conditionally calling retrieval, reasoning, and repair skills to fix violations (the 'Conceptual Rift'). Evaluated on the new Gen-Arena benchmark with the EGIP metric, SCOPE yields 0.60 EGIP and also performs strongly on WISE-V (0.907) and MindBench (0.61). Summary based on the abstract (full text not provided).

The paper tackles inverse imaging with Vision–Language Latent Diffusion Models by introducing a unified Euclidean–Wasserstein-2 gradient-flow that jointly samples the posterior and optimizes prompts in latent space. By pairing this flow with few-step latent text-to-image models, the method achieves low-NFE inference and avoids backprop through autoencoders, yielding state-of-the-art results on canonical inverse problems with much lower compute; summary based on the abstract only.

Lin and Liu show that mechanistic interpretability work often uses causal vocabulary (circuits, mediators, causal abstraction) while omitting explicit identification assumptions. Via a purposive audit of 10 papers and a two-coder check on 30 items, they document widespread substitution of validation metrics for identification and offer a five-step disclosure norm to make causal claims explicit. Submitted to NeurIPS 2026 (Position Track).

It Just Takes Two introduces a simple, theoretically grounded strategy for amortized neural posterior estimation that decouples representation learning from posterior modeling. The authors train a mean-pool Deep Set on sets of size ≤2 to produce an encoder that generalizes to arbitrary set sizes, then finetune an inference head on aggregated embeddings; this makes training cost essentially independent of deployment set size N. Across diverse benchmarks (scalar, image, multi-view 3D, molecular, high-dimensional generation) with N in the thousands, the method matches or outperforms baselines while using much less compute.

The Visa Business Card Companion Fare Offer in Spencer Burleigh’s Alaska Airlines (Atmos Rewards) account is set to expire on August 8, 2026 (about three months after the 2026-05-10 AwardWallet notice). AwardWallet reports the account was last updated 101 days ago and urges logging into AwardWallet to verify the coupon using provided auto-login and account-restore links.

The authors apply deep learning to infer global asteroseismic parameters (Δν, νmax, and for K2 also ΔΠ1) from short, one-month lightcurves to enable scalable analysis of TESS/K2 red-giant samples. Their model recovers Δν and νmax for ~50% of one-month Kepler/K2 cases but only ~23% for single-sector TESS; it produces ~200 reliable ΔΠ1 measures that match the Kepler Δν–ΔΠ1 sequence. (Summary based on the abstract; full text was not provided.)

The paper shows that language models (270M–27B, including Gemma 3, Qwen, Llama 3.1) encode tool selection as a linearly readable vector: adding per-tool mean-difference vectors reliably flips name-only single-turn tool choices and causes downstream JSON arguments to match the new schema. Causal attribution concentrates on one output-row and a few mid/late attention heads; base-model representations already carry tool identity (69–82% recoverable), while instruction tuning wires it to generation. Measurements are for single-turn fixed-menu settings; multi-turn transfer is noted as more fragile.

123D unifies multi-modal driving datasets by representing each sensor or annotation as a timestamped event stream, allowing flexible synchronization across heterogeneous formats. The authors merge eight real-world datasets (3,300 hours, 90,000 km) and a synthetic generator, perform systematic analyses of annotations and pose/calibration, and demonstrate cross-dataset 3D detection transfer and RL planning; the framework and tools are released open-source.

Karpathy recommends appending 'structure your response as HTML' to LLM prompts to render outputs in a browser, argues audio will be humans' preferred input while vision becomes AI's preferred output, and outlines a progression from text→markdown→HTML→interactive neural videos (eventually diffusion‑generated), urging improved multimodal inputs before BCIs.

The Garry's List Action Voter Guide, published May 11, 2026 by Garry Tan, Shaudi Fulp, and Forrest Liu, aggregates endorsements from local housing groups, labor unions, and civic reform organizations ahead of the California June 2 primary. The resource is presented as transparent, sourced, and searchable at garrysguide.org/elections and accepts suggested additions via email or X.

The paper addresses zero-shot human motion generation under very challenging spatiotemporal constraints by augmenting training-free diffusion noise optimization with retrieval guidance. It parses task constraints into groups (relational task parsing, powered by an LLM), retrieves reference motions for the hardest constraints, and forms a reward-guided mask to blend retrieved and random noise for improved diffusion initialization. The authors report this approach successfully handles tasks that prior methods struggle with, enabling more reliable constrained motion synthesis; accepted to CVPR 2026 (arXiv 2026-05-08).

CMR-EXTR converts free-text cardiac magnetic resonance (CMR) reports into auditable structured data with per-field confidence. The method employs a teacher–student distillation pipeline for fully offline inference and reduced annotation effort, plus an uncertainty model that blends distribution plausibility, sampling stability, and cross-field consistency to prioritize human review. It achieves 99.65% variable-level accuracy; code on GitHub; accepted to ISBI 2026.

The note proves that finite Gaussian surface area Γ implies existence of non-negative degree-k polynomials that ε-approximate indicator functions in L1 under the standard Gaussian, with k = ~O(Γ^2/ε^2). This adds a pointwise non-negativity guarantee (range [0,∞)), sits between plain L1-approximation and sandwiching polynomials, matches prior degree bounds up to constants, and targets smoothed positive-only learning. Only the abstract was available for this summary.

Reinforcement-learning for exponential-utility optimization in discounted MDPs: the paper derives two Q-value–style Bellman extensions whose operators are contractions in L_infty and sup-log/Thompson metrics, proves fixed-point structure and optimality of the induced greedy stationary policy among stationary policies, and presents two model-free algorithms — a two-timescale Q-learning with a.s. convergence and finite-time rates, and a one-timescale power-law method whose convergence is established via delicate local arguments. Full text on arXiv (abstract used).

Memelord’s May 11, 2026 newsletter from founder Jason “The Memelord” Levin claims the Memelord app has in‑app 'AGI for memes' and corrects a prior wrong download/link. The note links to the app, promotes a free trial at Memelord.com, and offers promo code MYTHOS for 50% off signups.

CA-SQL is a complexity-aware Text-to-SQL inference pipeline that scales exploration breadth by estimated task difficulty, employs evolutionary-search-inspired prompt seeding to elicit diverse candidates, and uses a novel voting method to pick final queries. On the Bird-Bench development set it reports 51.72% on the "challenging" tier (SOTA among in-context approaches with GPT-4o-mini), plus 61.06% execution accuracy and 68.77% Soft F1.

Imagined speech decoding: the authors recorded paired listened and imagined MEG from trained musicians and built a three-stage pipeline that maps imagined MEG to listened MEG (six mapping models), decodes words with a contrastive listened-only decoder (four embedding strategies), and applies the decoder to mapped imagined data from held-out subjects. Proof-of-concept results show significant above-chance decoding and scalability with more training data; only the abstract was available for this summary.

GRAPHLCP tackles conformal prediction for graph neural networks by addressing failures of embedding-space localization on graphs. It combines feature-aware densification to reduce locality bias in sparse graphs with a Personalized PageRank kernel to capture structural proximity for anchor sampling and calibration weighting. The approach yields finite-sample marginal coverage and empirically better test conditional coverage and smaller prediction sets on several regression and classification benchmarks compared to prior embedding-only localization methods.

The paper 'The Memory Curse' (Liu et al., 2026) shows that expanding LLMs' context windows often erodes cooperation in multi-agent social dilemmas: across 7 models and 4 games over 500 rounds cooperation fell in 18 of 28 model–game settings. Analyses of 378,000 reasoning traces implicate loss of forward-looking intent; targeted LoRA fine-tuning, memory sanitization, and CoT ablations partly restore cooperation. (Based on the abstract.)

NoiseGate reframes per-latent timestep selection in joint video–action world-action models as a learnable information-gating policy: by adjusting each predicted latent frame's noise level, a Gating Policy Network controls its Key/Value reliability for action generation. The approach (independent per-latent timestep sampling plus task‑reward optimization) improves RoboTwin random-scene manipulation performance. Summary based on the paper's abstract; full text not available here.