Leaderboards, eval harnesses and ablations — the contested business of deciding which model is actually better.
arXiv:2609.13990v1 Announce Type: cross Abstract: Diffusion-based generative models (DMs) have found applications in control problems, and in particular robotics, where the DMs enable exploration of...
VisInteract introduces a new paradigm for Text-to-Visualization that treats imperfect user queries as a dynamic, interaction-driven problem, requiring the system to recover true intent through multi-turn dialogue. The authors present VisInteract-Bench, the first benchmark for interactive Text-to-Vis, featuring controlled imperfection injection, a realistic user agent, and dual-perspective automated evaluation. They also propose Vis-MCTS, an enhanced Monte Carlo Tree Search algorithm that incorporates progressive widening, cross-rollout information sharing, and dimension-aware reward decomposition, achieving significant performance gains over existing baselines.
Stellar Colosseum is a model‑agnostic harness designed to improve long‑horizon research in mathematics and theoretical computer science by allocating inference across multiple agents. It explores alternative strategies before constructing proofs, uses a readiness gate to decide when a route is mature enough to decompose, represents proof plans as interdependent subproblems, and routes verifier findings back to the relevant part of the argument. The workflow generates candidates in parallel, attacks them with targeted falsification, and combines candidates and critiques into a single research artifact through overlapping random‑sample tree aggregation, and has been integrated into Google Antigravity's Teamwork framework as the Long Proof pattern. Demonstrations show that, when paired with Gemini 3.1 Pro, Stellar Colosseum achieves 71.0% accuracy on the TCS‑Bench theorem‑proving benchmark and solves 218 of 222 Codeforces problems.
The paper introduces the Compute-Value Audit (CVA), a sequential framework that evaluates whether extra sampling during test‑time scaling for video world models actually yields a net benefit after accounting for the compute cost of generation and verification. On 192 Physics‑IQ scenes, increasing the sample pool from 4 to 16 candidates improves oracle quality by +9.23 IQ, yet common metrics such as Flow, Cycle, and VideoReward fail to reliably recover this headroom, and adaptive‑depth policies recover only 42‑69% of the potential gain. Only a few specific interventions—anchor‑explorer in a sparse PRM800K setting, MMLU‑Pro exposing a predictive‑state gap, and a privileged paired‑future upper bound—successfully pass all CVA stages, indicating that sampling headroom is valuable only when it can be converted into a reliable decision that survives the full compute charge.
The paper introduces the Variational Incompressible Optimal Transport (VIOT) operator, a generative neural operator that predicts divergence‑free velocity fields for incompressible density transport. VIOT combines a stream‑function representation, a regularized transport objective, and a Fourier Neural Operator backbone to amortize the solve across new source‑target pairs and grid resolutions. Experiments on 2D and 3D benchmarks show that VIOT produces full transport trajectories in seconds, achieving roughly a $10^4 imes$ speedup over per‑instance baselines that require hours of optimization.
The paper introduces ETHEREAL, the first accelerator for event-driven graph neural networks (EV‑GNNs) that can handle 640×480 resolution inputs. It achieves this through a neighbor‑parallel spline convolution engine and a 2D/3D‑split memory hierarchy that includes a novel region‑of‑interest spatiotemporal caching mechanism. Measurements show end‑to‑end inference latency of 25.6 µs and energy consumption of 1.7 µJ per event on state‑of‑the‑art workloads.
The paper introduces SCALAR, a query‑conditioned spherical centroid aggregator that assigns relevance‑based weights to each available modality before computing a spherical centroid. SCALAR supports arbitrary modality subsets, is trained with rank‑8 LoRA adapters on masked views, and achieves positive aggregation gains on four of five benchmarks, outperforming prior symmetric aggregators. With only 4.8 million trainable parameters, SCALAR attains the highest text‑to‑video R@1 on three benchmarks and surpasses the released GRAM checkpoint under test‑time modality dropout by 3.2 to 10.9 R@1.
MambaMPD is a new segmentation framework that leverages Vision Mamba models for marine pollution detection in remote‑sensing imagery. It introduces two structural priors—Frequency‑Aware Augmentation (FAA) and multi‑scale Edge‑Guided Attention (EGA)—to better capture low‑contrast, fragmented pollution patterns and sharpen boundaries. Experiments on the MADOS and M4D datasets show that MambaMPD outperforms existing methods in mIoU while using far less computation than foundation‑model approaches.
arXiv:2604.19657v2 Announce Type: replace-cross Abstract: AI agents promise to serve as general-purpose personal assistants for their users, which requires them to have access to private user data (e...
The paper introduces self‑orchestrating language models that annotate semantic dependence—identifying which tokens rely on others—to guide efficient inference. By leveraging these annotations, the authors design runtimes that parallelize autoregressive decoding, evict intermediate context, or determine denoising orders, achieving Pareto‑optimal quality‑efficiency trade‑offs. Three systems—PASTA, TIP, and Planned Diffusion—demonstrate these techniques for parallel decoding, memory‑efficient reasoning, and efficient discrete diffusion, respectively.
GRIN+ is a new machine unlearning framework that targets fast and precise data erasure in imbalanced medical datasets. It separates unlearning‑specific knowledge from general representations by analyzing gradient contributions of forget and retain sets, introduces a class‑adaptive influence scoring to counter gradient dominance, and uses a direction‑constrained update to protect essential clinical knowledge. Benchmarks on skin cancer, brain tumor, and breast ultrasound data show that GRIN+ balances privacy, efficiency, and utility, achieving high diagnostic accuracy and faster runtime than existing methods.
Habitat-GS is a navigation-focused embodied AI simulator that extends Habitat‑Sim by incorporating 3D Gaussian Splatting (3DGS) for real‑time photorealistic rendering and scalable asset import. It introduces a Gaussian avatar module that represents dynamic humans as both photorealistic visual entities and navigation obstacles, enabling agents to learn human‑aware behaviors. Experiments show that agents trained on 3DGS scenes generalize better across domains and that the avatar system supports effective human‑aware navigation while maintaining system scalability.
EEG-Xplain introduces a unified attribution framework to interpret EEG foundation models such as BIOT, LaBraM, and EEGMamba. The framework combines gradient, perturbation, and activation-based methods to analyze model behavior across spatial, temporal, and frequency dimensions, identifying critical channels, decision-relevant signal segments, and contributions of canonical EEG rhythms. It evaluates explanation reliability with population-level metrics and uses large language models to convert structured attributions into natural-language reports, demonstrating consistency with known neurophysiological markers on benchmark datasets.
The paper demonstrates that a single example from the BBQ fairness benchmark can dramatically improve a model’s performance, raising accuracy from 79.9% to 92.9% with Group Relative Policy Optimization and to 99.0% with one-shot in-context learning. This effect is consistent across different model families and is driven by the model’s reasoning traces, which adopt a category‑agnostic "missing evidence" pattern. The authors argue that BBQ-style multiple‑choice abstention tests capture only a single structural cue and therefore do not guarantee true fairness, calling for broader evaluation suites.
The paper investigates how large language models handle domain-specific jargon, comparing a general-purpose Llama‑3.1 with a version fine‑tuned on medical data. Two new medical jargon benchmarks reveal that the general model actually outperforms the fine‑tuned variant, and interpretability tools show the fine‑tuned model over‑emphasizes a few components linked to jargon predictions. Reweighting these components narrows the performance gap, and some jargon‑sensitive components also aid materials‑science tasks, indicating a partially domain‑agnostic representation of specialized terminology.
The paper introduces a reverse spatio‑temporal disease progression model that reconstructs unobserved healthier anatomy from later diseased scans. It employs a two‑stage architecture: a frozen 3D vector‑quantised autoencoder creates a discrete latent space, and a Neural ODE learns continuous‑time dynamics, with a recurrent encoder initializing the latent state from reverse‑ordered observations. Experiments on a synthetic Morpho‑MNIST benchmark and longitudinal Alzheimer’s MRIs show the model can recover unseen prior states and outperform baseline methods in predicting healthy trajectories.
The study investigates how fine‑tuning a large language model on a single task‑language pair influences performance on other task‑language pairs. Using LoRA fine‑tuning across multiple open‑weight LLM families, 11 languages, and four benchmarks, the authors decompose transfer into matched‑task, cross‑task, and cross‑task cross‑language regimes. They find that while single‑source fine‑tuning generally improves performance, the gains are highly asymmetric, with matched‑task cross‑language transfer being most effective and driven mainly by the target language rather than model architecture.
SDUs DAISY is a factual knowledge benchmark focused on Danish cultural heritage, drawing topics from the Danish Culture Canon 2006. The dataset contains 741 manually verified closed‑ended question‑answer pairs generated by querying Wikipedia pages for each canon artifact, covering a wide historical span from 1300 BCE to contemporary pop music, design, and architecture. Baseline tests with state‑of‑the‑art language models show very low performance (best BLEU 0.17, F1 0.27), highlighting the benchmark’s difficulty and the need for improved cultural knowledge in AI systems.
The paper introduces NegCue, a large-scale dataset of 1.8 million samples that includes single-word, multi-word, and affixal negation cues, totaling over 200 unique forms. The authors pre-train encoder-only language models and large language models on this dataset to study how different negation types influence understanding. Experiments on five downstream benchmarks reveal that affixal negations provide the most significant performance gains, whereas single-word negations yield modest improvements, and that additional pre-training benefits both model types.
The paper introduces CarryOnBench, an interactive benchmark that tests whether large language models can revise their interpretation of user intent and recover utility while staying safe in multi‑turn conversations. Using 398 harmful‑looking queries with benign intents, the benchmark simulates 5,970 conversations across 14 models, evaluating both intent‑aligned utility and safety with a new metric called Ben‑Util. Results show that models often withhold information due to misinterpretation, but most can recover with clarifications, revealing failure modes such as unsafe and redundant recovery that single‑turn tests miss.