Leaderboards, eval harnesses and ablations — the contested business of deciding which model is actually better.
arXiv:2610.27323v1 Announce Type: cross Abstract: Vision-language models (VLMs) can infer sensitive information such as geolocation from subtle visual cues, creating privacy risks when responses disc...
arXiv:2610.27439v1 Announce Type: cross Abstract: Solving reasoning tasks supported on graphs typically requires providing a Language Reasoning Model (LRM) with a serialized description of the graph....
The paper proposes a training‑free continual learning approach for multimodal time‑series forecasting by evolving a textual guidance block that is appended to a frozen large language model’s prompt. On a benchmark combining weekly search indices, daily currencies, and daily commodities with web‑retrieved news, the evolved 900‑word guidance outperforms existing time‑series foundation models and transfers to newer models without loss, while requiring only monetary cost rather than GPU days. The authors also show that compiling the guidance into executable code recovers most of the gain deterministically and that the advantage grows at short horizons, with news text being distilled into policy during learning.
arXiv:2610.27533v1 Announce Type: cross Abstract: Machine learning benchmarks act as guideposts for model development and spur the creation of training datasets intended to improve benchmark scores....
arXiv:2610.27566v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly used within everyday applications where their responses have significant downstream impacts. For respon...
arXiv:2507.21035v4 Announce Type: replace Abstract: Gene expression analysis holds the key to many biomedical discoveries, yet extracting insights from raw transcriptomic data remains challenging due...
arXiv:2610.28631v1 Announce Type: new Abstract: Current world models for explicit world simulation typically follow two distinct pathways: (1) video generative models (e.g., Sora), which simulate rea...
arXiv:2508.11850v3 Announce Type: replace Abstract: Integer programming (IP) is central to many combinatorial optimization tasks but remains challenging due to its NP-hard nature. A practical way to...
arXiv:2610.24452v2 Announce Type: replace-cross Abstract: We introduce avatar referring segmentation, which selects regions of animatable Gaussian avatars from language queries. A query can also spec...
The paper examines how humans and multimodal large language models (MLLMs) generate associative descriptions for abstract visual stimuli that become harder to name. Using two datasets—SilCC, which increases difficulty by combining abstract shapes, and PixCC, which adds visual noise—the study measures originality and effectiveness of the generated descriptions. Findings show that humans produce more diverse, semantically distant, and novel descriptions than MLLMs, and that as difficulty rises, human effectiveness drops in SilCC while they shift toward low-level visual descriptions in PixCC.
EGI‑BENCH is a unified benchmark suite for evaluating large language model agents in enterprise settings, consolidating eight existing benchmarks (CRMARENA, HERB, PERSONABENCH, USERBENCH, MCP‑UNIVERSE, MCPEVAL, BENCHFORCE, LOCOBENCH) under a single command‑line interface with two runtime‑budgeted modes. It introduces the Enterprise General Intelligence (EGI) Score, a six‑dimensional framework (M1‑M6) that aggregates fine‑grained sub‑metrics from each benchmark rather than simple averages, providing a more nuanced view of model capabilities. The paper reports results for fourteen LLMs across GPT, Claude, and Gemini families, showing varied strengths across task outcome, planning, retrieval, tool actioning, safety, and engineering quality.
TEX is a new test‑time scaling framework that uses execution‑driven cross‑validation to generate and refine tests for language‑model agents. By running agents in parallel and cross‑executing each test against every patch, TEX builds a $K imes K$ feedback matrix that guides objective refinement without sharing verbose traces. On the SWT‑Bench Verified benchmark, TEX reaches 86.4% accuracy—setting a new state‑of‑the‑art and improving over single‑agent baselines by 9.0%—and even random selections from its outputs outperform the previous best score.
The paper introduces REM (Reusable Experience Memory), a framework that manages the lifecycle of experience for large language model agents performing long-horizon coding tasks. REM consolidates completed trajectories into compact, phase-structured documents, injects them at task start, and evicts them after a single exposure, making the memory system responsive to the agent’s needs. Experiments on repository-level bug fixing show REM outperforms state‑of‑the‑art memory systems in retrieval accuracy and task resolution while reducing cost.
Semantic Observability for Autonomous Systems Diagnostics proposes an autonomous multi‑agent framework that consolidates heterogeneous observability logs into a dynamic, queryable Property Graph Database. By modeling infrastructure topologies, software dependencies, and event sequences as attributed nodes and directed relations, the system applies causal graph algorithms and temporal back‑traversal to isolate anomaly blast radii. A Large Language Model then interprets the pruned subgraph to produce code‑level root cause analyses and actionable remediations, aiming to reduce MTTR and LLM token consumption while preventing hallucinations.
The paper introduces EmbGA, an embedding‑space genetic algorithm designed to uncover hidden concepts in text‑to‑image models that are not explicitly mentioned in prompts. Unlike existing prompt optimizers, EmbGA operates directly in token embedding space, using crossover and mutation to generate valid token sequences that align with unseen visual concepts. Experiments on two benchmarks show that EmbGA outperforms six other optimizers in alignment and token grounding, especially with larger exploration budgets, while also being faster than language‑model‑guided search.
World-Action Models (WAMs) jointly predict future world states and robot actions for diverse manipulation tasks, but they can suffer from Stage Confusion—generating plausible yet misdirected rollouts that drift to the wrong task stage. TaskWAM introduces a closed-loop, stage-aware framework using Mode Tokens and State Tokens to represent fine-grained task requirements and execution progress, steering generation toward the intended future. The authors also present StageBench, a multi-stage manipulation benchmark, and show through extensive simulation and real-world experiments that TaskWAM reduces confusion, matches state-of-the-art performance, and transfers effectively across various robot embodiments.
arXiv:2604. 07667v2 Announce Type: replace Abstract: Multi-agent debate improves LLM reasoning, yet agreement is not evidence of correctness.
The paper introduces a curriculum learning strategy for generative flexible docking, a task that predicts protein‑ligand complex configurations from an apo protein structure and a ligand SMILES string. By gradually incorporating examples with large conformational changes during training, the authors enhance an existing generative model, achieving a 46.52% top‑1 success rate on the PDBBind ESMFold apo benchmark and outperforming prior baselines. The approach leverages diffusion and flow‑based generative models to model the posterior distribution over ligand poses and protein conformations conditioned on the apo input.
arXiv:2610. 28355v1 Announce Type: new Abstract: In speculative decoding, a lightweight draft model proposes tokens that the target model verifies in parallel, speeding up generation without changing the target distribution.
The paper investigates how CLIP-style dual encoders sometimes increase similarity scores when a text description is extended with a plausible but incorrect detail, a phenomenon termed "half-truths." On the COCO dataset, CLIP correctly prefers the shorter, accurate description only 40.6% of the time, dropping to 32.9% when the added detail is a relation. The authors attribute this to weak supervision on caption components and propose CS-CLIP, a component-supervised fine-tuning approach that raises half-truth accuracy to 69.3% and improves compositional benchmark performance by 5.7 points.