As enterprise agent tool catalogs scale from 10 to 110 agents, routing accuracy drops 16--23 percentage points on under-specified requests. An oracle analysis identifies retrieval and confusion gaps, with embedding-based shortlisting recovering +10--11pp F1. A human-annotated study of 1,435 utterances confirms real-world recovery of +10--17pp despite lower absolute performance.
This paper proposes SD-ZFS, a deep reinforcement learning framework adapted from S2V-DQN, to solve the NP-hard minimum zero-forcing set problem on undirected graphs. The framework demonstrates strong performance compared to optimal solutions and greedy heuristics, showing effective generalization, scalability, and transfer across diverse graph structures.
A new convex quasilinearization method, LiL-Q, solves nonlinear PDEs by reducing them to linear subproblems using physics-informed neural networks. LiL-Q converges in single-digit iterations across seven benchmarks, achieving machine precision when the exact solution lies in the trial space, and requires up to two orders of magnitude fewer parameters than standard PINN solvers.
A study of 14,727 security and privacy prompts from 3.2M real-world user-LLM conversations identifies nine categories of S&P queries. Commercial LLMs outperform open models, with GPT 5.5 providing good responses on 98% of prompts versus Llama 4 at 47%, though some commercial models produce contradictory responses across runs.
MGUP introduces a selective update mechanism that applies larger step-sizes to a fixed proportion of parameters in stochastic optimization, while using smaller, non-zero step-sizes for the rest. It integrates seamlessly with optimizers like AdamW, Lion, and Muon, providing theoretical convergence guarantees for MGUP-AdamW and demonstrating superior or more stable performance in training large language models and MAE pretraining tasks.
SPHERE-JEPA introduces deterministic statistical regularizers on the hypersphere, replacing stochastic sliced methods with analytically integrated objectives like MMD, KSD, and KL divergence. Rotationally invariant kernels based on heat and bandlimited filters ensure spatial bias-free learning, with empirical results showing improved convergence and performance on ImageNet and Galaxy10, and superior instance separation in procedural texture retrieval using KL divergence.
TUNEAHEAD is a lightweight framework that predicts fine-tuning performance using meta-feature vectors from dataset descriptors and short probe runs. It outperforms baselines like Early-Stop Extrapolation and ProxyLM, achieving an RMSE of 1.47 percentage points and 95.1% of predictions within ±3 percentage points of true scores on 370 held-out runs.
A confusion-aware difficulty score is introduced within the Transfer Teacher framework to improve model interpretability and data efficiency. Evaluations on CIFAR-10 show that confusion-aware curriculum ordering outperforms random ordering by up to 8.7% at 20% data, demonstrating consistent data-efficiency gains. However, curriculum or anti-curriculum ordering does not improve accuracy over standard training at full data, indicating that scoring function improvements alone are insufficient to overcome curriculum learning failure modes.
The paper introduces 'No-Free-Fairness' theorems that prove three fundamental limits in learning systems. These include inherent fairness-cost trade-offs, unavoidable subgroup disparity in finite samples, and model expressivity constraints that prevent fairness regardless of data. The results show fairness is constrained by problem structure, data limits, and model capacity, not just biased data.
A study of 14,727 security and privacy prompts from 3.2M real-world user-LLM conversations identifies nine categories of S&P questions. Thematic analysis and response testing show commercial LLMs outperform open models, with GPT 5.5 providing good responses on 98% of prompts versus Llama 4 at 47%, though some commercial models produce inconsistent responses across runs.
A study evaluated several AI systems on ten research-level mathematics problems created by prominent mathematicians. The results include AI-generated solutions, human solutions, and referee reports, offering a detailed assessment of AI performance in solving advanced mathematical problems.
Language models of GPT-2 size cannot independently discover zero during testing, regardless of pretraining. However, performance improves significantly with training on tens to hundreds of zero examples, and language pretraining reduces required examples by about 50%.
A new framework uses prompt perturbation to identify and filter structurally inconsistent pairwise comparisons in large language model evaluations. By incorporating graph-level consistency checks before ranking aggregation, the method reduces cyclic preferences and improves the reliability of LLM rankings.
We present a framework for evaluating agentic skills by constructing realistic tasks and assessing skill utility through task execution. Applied to 500 real-world skills, it generates 1,000 tasks and scoring rubrics, evaluating 19 agent-model configurations across proprietary and open-source models. Results show significant variation in instruction adherence and performance gains, with skills substantially altering model behavior compared to no-skill setups.
NNN decoding selects documents as a joint set that jointly reconstructs the query embedding via a sparse non-negative linear combination. It strictly extends dense retrieval by handling queries that dense retrieval fails on, especially in corpora with correlated documents, and achieves superior performance through end-to-end training of embeddings.
Glimmer 1 is a 10,000-parameter language model trained on 500K tokens from FineWeb-Edu. It features a 512-token context window, a standard Llama architecture with 16 hidden dimensions, 2 layers, 4 attention heads, and 1 KV head using GQA, and is available on Hugging Face.
A measurement study finds that 26 semantic post-hoc operators do not improve held-out accuracy over Best-of-N in frozen small code models. While two operators—expression-layer recovery and adaptive consensus early-stop—offer benefits in compute efficiency or program recovery, none outperform BoN in accuracy. The results highlight systemic limitations in error detection and coverage, suggesting that model harnesses and error coverage must be improved before post-hoc reasoning is considered.
MA-SBI introduces a calibration-free simulation-based inference framework that uses side-channel text, like regime labels or instructions, to correct for simulator misspecification. It employs a learned corrector to apply observation-space shifts before posterior inference, without needing ground-truth parameter pairs or retraining. On hide-the-calibration benchmarks, MA-SBI matches the oracle posterior with text alone, outperforming RoPE under limited data, and shows robustness on real-world epidemiological and cognitive-science datasets.
Analysis of 56,800 AI conference papers shows documentation practices improved from 2014 to 2024. Papers sharing both code and data increased from 11% to 64%, and estimated reproducibility rose from 28% to 64%. These improvements predate formal reproducibility checklists, indicating a broader shift toward open science.
Public AI evaluation archives show that a single terminal result can arise from two distinct pre-terminal histories, with estimated times to reach 95% of performance ceilings at 23.03 or 75.13. A candidate selection-aware frontier model fails synthetic recovery and uncertainty calibration, and is rejected by fixed audit gates. An archive-and-adjudication protocol verifies timing boundaries and falsifies unsupported frontier claims.