Leaderboards, eval harnesses and ablations — the contested business of deciding which model is actually better.
arXiv:2609.37858v1 Announce Type: new Abstract: Many localized large language model (LLM) unlearning methods select a small parameter subset from a localization signal and keep it fixed during optimi...
The paper investigates patch‑wise supervision for detecting AI‑generated images, proposing a shared backbone that classifies explicit crops with individual losses and averages patch probabilities only during inference. This approach eliminates the need for handcrafted residual filtering or learned image‑level fusion modules. Experiments across single‑patch selection, multiple generator collections, and four CNN and Transformer backbones show that patch‑wise variants outperform whole‑image counterparts on the GenImage dataset, while also exploring factors such as supervision granularity, source resolution, crop size, and inference coverage.
Large Language Model judges are commonly used to rank texts via pairwise comparison, with reliability traditionally measured by position bias, transitivity, and pairwise agreement. This paper argues that these proxies are misleading because they are dominated by close‑rank‑gap pairs, which contribute little to the overall ranking, while far‑gap pairs carry the true ranking signal. Experiments on simulations and human‑rated corpora show weak correlation between the proxies and actual ranking accuracy, suggesting judges should be evaluated using rank‑gap‑conditional metrics against human rankings.
The paper introduces SCSP, a training‑free framework that improves long‑context embeddings by selectively pooling informative tokens. SCSP partitions documents into sentence‑aware chunks, adds a semantic compression prompt to each chunk, and uses prompt‑isolated attention masks to estimate token importance. The selected tokens’ intermediate‑layer representations are aggregated to form the final embedding, yielding consistent performance gains across zero‑shot and fine‑tuned models on long‑context benchmarks.
The paper introduces LongHarness Bench, a new benchmark designed to evaluate both the effectiveness and efficiency of language model harnesses for long-context reasoning. It features tasks that require diverse retrieval strategies—such as lexical search and semantic matching—and strategic, adaptive reasoning over global and local context, with only a small subset of the context being useful at each step. Evaluations across multiple state‑of‑the‑art models and harnesses show that even strong combinations achieve only 68% macro‑average accuracy, and that the same model can vary markedly in efficiency depending on the harness used.
The paper introduces RS-OPSD, a reliable privileged on-policy self-distillation framework designed for ultra‑high‑resolution remote sensing visual question answering. It leverages a new dataset, GeoEvidence‑6K, and a human‑feedback guided skill refinement process to provide explicit question‑relevant evidence. By incorporating context‑preserving visual privilege and correctness‑aligned distillation, RS‑OPSD achieves state‑of‑the‑art performance on several benchmarks without requiring additional visual search or tool calls at inference time.
AdaTutoRank introduces a setwise document reranker that uses Adaptive Tutoring Optimization (ATO) to provide graded supervision across nine rubric dimensions. By generating hint‑based silver labels, reinforcement rewards, and distillation cues tailored to each rollout’s quality, the method improves credit assignment for individual documents within a set. Experiments on ten benchmarks covering Retrieval‑Augmented Generation (RAG), deep research, and setwise evaluation show that AdaTutoRank achieves superior overall performance while reducing the number of retrieval calls.
arXiv:2502.04492v3 Announce Type: replace Abstract: The advancement of LLMs and their accessibility have triggered renewed interest in multi-agent reinforcement learning as robust and adaptive framew...
arXiv:2601.03418v3 Announce Type: replace Abstract: Trustworthy clinical summarization requires every claim to be traceable to its evidence, yet existing attribution often resolves only to the senten...
The paper introduces RED (Reconstruction Evolution Dynamics), a new framework for detecting AI-generated images that leverages the evolution of intermediate reconstruction stages rather than relying solely on static representations or endpoint discrepancies. RED uses a frozen multiscale VQ‑VAE and a frozen CLIP encoder to capture a reconstruction trajectory, then learns image‑adaptive stage weights from token negative log‑likelihoods provided by a frozen VAR model. Experiments on six benchmarks show RED achieves the highest average accuracy (92.5%) and precision (97.5%) among evaluated methods, and it remains robust to common image degradations.
arXiv:2609.37853v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly deployed as social agents, yet credible human-like interaction requires more than fluent responses or per...
arXiv:2609.36440v1 Announce Type: new Abstract: Large vision-language models (LVLMs) have recently achieved remarkable progress across multimodal tasks, yet object hallucination remains a persistent...
arXiv:2609.38180v1 Announce Type: new Abstract: Existing 3D part decomposition methods do not necessarily partition the original shape into non-overlapping parts that collectively cover the entire sh...
arXiv:2609.35910v1 Announce Type: new Abstract: Adaptive perception spends extra computation on inputs where perception is expected to improve. When perception feeds a downstream decision system, a b...
The paper introduces TMLN (Trajectory-Modulatory Landscape Navigation), a method that treats continual learning as an optimal control problem on a curved loss landscape. It uses a diagonal empirical Fisher Information Matrix to approximate a local Riemannian manifold and dynamically modulates a preconditioner based on the network’s historical parameter trajectory. This trajectory‑based preconditioning is integrated into gradient updates to protect important parameter directions without adding explicit penalties, and experiments on class‑ and domain‑incremental benchmarks show a significant reduction in the loss barrier between tasks.
arXiv:2609.36064v1 Announce Type: cross Abstract: Foundation models are powerful generators, but many engineering domains require structured representations that general-purpose systems handle poorly...
arXiv:2609.36144v1 Announce Type: new Abstract: Electronic health records provide irregular observations of latent patient states that evolve continuously over time. Recent autoregressive models cond...
arXiv:2609.36683v1 Announce Type: new Abstract: Molecular optimization is inherently iterative: a candidate is proposed, evaluated against several objectives, and revised while preserving a relations...
arXiv:2609.36881v1 Announce Type: new Abstract: Causal foundation models (CFMs) pre-trained on data generated from various structural causal models (SCMs) have been proposed for estimating causal eff...
arXiv:2609.36890v1 Announce Type: new Abstract: In scientific machine learning, physical fields governed by partial differential equations exhibit low-rank structure and scale invariance. When solvin...