Inference & Efficiency A

The authors present a convolutional variational autoencoder for crypto implied-volatility surfaces, paired with a predictor combining it with a quadratic smile re-fit via a deterministic per-tenor routing rule. Trained on 6,034 hourly Binance BTC and ETH option surfaces (May-October 2023), it achieves hidden-cell completion RMSE of 0.94-1.56 vol points. At 50% masking the hybrid reaches 0.83 vol points versus 7.00 for the smile re-fit alone, an eightfold reduction at no extra inference cost.

On its developer blog, NVIDIA explains advanced fusion-kernel techniques aimed at boosting training throughput for Mixture-of-Experts (MoE) models. Noting that MoE has rapidly become a foundational component of modern large-scale AI systems, the post outlines kernel-level optimizations for more efficient training.

Theory of mind (ToM), ascribing mental states to others for prediction and inference, is widely assumed essential for human-machine integration. Existing AI-ToM models address how to mentalize but leave when largely unaddressed. The paper asks under what situational and agent-level conditions ToM engagement is causally warranted in conflict, presenting a structural causal model as a directed acyclic graph that treats ToM as a mechanism activated by conditions rather than an always-on capacity.

This paper studies reasoning with a Code Interpreter (CI) in LLMs from two angles: extrinsic properties (crucial tokens) and intrinsic properties (code-specific cognitive behaviors). It reports that stronger CI reasoning models show more crucial tokens and behaviors—especially verification, backtracking, and backward chaining—and explores leveraging these at inference and training time. Summarized neutrally from the abstract.

Time-series foundation models transfer well given a history window, but true cold-start items with no prior observations violate that. The authors propose RAID (Retrieval-Augmented Iterative Diffusion), replacing history-based correlation with metadata-driven semantic retrieval and graph-conditioned diffusion. It maps metadata into a shared semantic space via a frozen multilingual embedding model, builds an inductive retrieval graph for unseen items, and refines a forecast from neighbors.

Simulation-based inference (SBI) is often hindered by simulator misspecification, the mismatch between simulated and real observations. The recent robust method RoPE uses optimal transport between learned representations but needs ground-truth calibration pairs unavailable where SBI is needed. Practitioners instead have unstructured side-information such as regime labels, instruction text, and policy bulletins. The authors propose Misspecification-Aware SBI (MA-SBI) to exploit this guidance.

The paper presents LESS, a training-free, model-agnostic adaptive sampler for diffusion LLMs that frames token commitment as an online stopping problem. Its mutual-stability rule unmasks a position only when its top-1 prediction is confident, persists across recent steps, and is distributionally stable (top-K inter-step JS divergence). It is evaluated on Dream-7B, LLaDA-8B, and LLaDA-1.5-8B. Summarized neutrally from the abstract.

The paper addresses spatial question answering for service robots traversing long egocentric routes, returning metric coordinates that downstream navigation can act on for queries like 'where can I find a dry cleaner on the way back home?' Prior approaches rely on closed-source models such as GPT-4o, which robots cannot reliably depend on due to network instability, latency, and deployment cost. The authors propose Binary Tracking, an open-source vision-language approach that can run onboard.

An arXiv paper addresses the speed-quality trade-off and error propagation in revocable decoding for diffusion LLMs (dLLMs). It proposes following a latent 'roadmap' guided by anchor tokens to mitigate failures arising in mixed-quality contexts during parallel generation. Neutral, abstract-based summary.

An arXiv paper introduces Expert Tying, an architectural change that shares expert parameters across consecutive transformer layers while keeping independent layer-wise routing and attention, aiming to cut Mixture-of-Experts memory cost. Summarized neutrally from the abstract.

The paper introduces GIST-CMTF, which augments Causal Minimal Tool Filtering with goal-state inference for tool-augmented LLM agents. It addresses wrong-goal execution, where ambiguous requests such as "handle my appointment" map to multiple goals and an agent may follow a valid causal tool path toward an unintended objective.

An arXiv paper presents an LLM-based visual programming copilot for aerospace geometric design tasks, using a visual-programming variant of the ReAct methodology. Summarized neutrally from the abstract; claims are the authors' and not independently verified.

The paper proposes a progressive, physics-guided multi-scale vibration-processing pipeline for bearing fault diagnosis, using a kinematics-derived descriptor for real-time screening and fault-adaptive segmentation. Reported figures reflect the abstract and are not independently verified.

This arXiv paper proposes DoubtProbe, a dual-branch inference-time framework for black-box jailbreak defense in LLMs. The authors observe that many jailbreaks do not remove the harmful goal but reorganize the information needed to express it, evading safety alignment while remaining recoverable during generation. DoubtProbe combines structural verification and semantic auditing to counter this.

Sakana AI has launched Sakana Marlin, its first commercial product: an autonomous research assistant for business. Given a research theme, it works autonomously for up to about eight hours—forming hypotheses, gathering and verifying information—then outputs structured summary slides and a report spanning dozens of pages. Built on the firm's long-horizon reasoning technology, it aims to act as a 'virtual CSO,' is self-serve, and available same day, with plans from free pay-per-use to Enterprise.

An essay on the future of voice assistants like Siri, arguing that on-device or 'private' inference alone does not fully protect user privacy and that stronger guarantees are needed beyond encryption and local processing.

NVIDIA reports leading agentic coding performance on the first benchmark dedicated to agentic AI, per its developer blog. The result highlights its inference stack and GPU infrastructure as a platform for autonomous coding agents.

AdaSR moves beyond the read-then-think paradigm by letting reasoning models reason incrementally as input streams in. It uses a hierarchical relative policy optimization scheme to train streaming reasoning.

HumP-KD is a hybrid, uncertainty-aware multi-stage progressive knowledge distillation framework for fire classification. It targets models that are simultaneously accurate and efficient for real-time use.

LLMs serve as execution engines for agentic systems yet still consume context through a sequential text interface, mismatching modern structured workflows with independent parallel branches. The paper explores synthesizing such parallel branches directly in latent space.

This work targets knowledge editing that updates selected facts while preserving nearby behavior in a memory-assisted setting. It proposes route-specialized dual adapters that decide when to write and when to suppress edits.

Time-stamped interaction logs objectively record digital app usage, but their granularity and noise obscure meaningful insights into work. The paper proposes abstracting cross-domain action sequences into interpretable workflows.

Through computational analysis, this paper argues that the three movements of Beethoven's Moonlight Sonata (Op. 27 No. 2) instantiate three distinct machine learning architectures by structural correspondence rather than mere analogy.

Post-training INT8 quantization of diffusion transformers is often slower than FP8/NF4 on consumer Ampere GPUs. The paper presents a fused INT8 GEMM kernel for Ideogram 4.0 that realizes native INT8 speedups.

The paper studies whether transformer-based neural operators for PDE solution operators can generalize zero-shot to larger spatial domains than seen in training.

Giving LLMs direct control over costly, irreversible experiments invites unsafe exploration, while discarding their creativity sacrifices optimization. CARE controls LLM-generated policies through auditable review of evidence in scientific experimentation.