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.
We introduce Precision-Cache Hit Ratio (P-CHR) AUC and Calibration Retention Rate (CRR) to address the calibration gap in semantic caching. These metrics evaluate precision across cache utilization levels and measure how offline ranking quality persists in deployment. Our analysis shows the gap is driven by training objectives, not data scale, and post-hoc calibration only partially resolves it.
Multi-Agent Transactive Memory (MATM) enables population-level storage and retrieval of agent-generated trajectories. It allows producer agents to share procedural knowledge with consumer agents, improving task performance and reducing interaction steps in interactive environments like ALFWorld and WebArena without coordination or joint training.
A study measures tool-intent stabilization in Streaming RAG, defining when speculative tool queries converge to correct answers. On the CRAG benchmark, 73.9% of queries allow substantial latency hiding, with early stabilization observed in questions with verbatim retrievable evidence. Question type significantly predicts early versus late stabilization, informing when speculative triggers are effective.
CATCH-ME introduces the first large-scale, multilingual dataset of contextually annotated, multi-turn counterspeech dialogues targeting hate and misinformation. The dataset covers five languages and focuses on seven marginalized groups, with dialogues grounded in verified fact-checking sources and including document- and chunk-level span annotations for RAG systems.
A new metric, Cross-dataset AUC (Cross-AUC), addresses limitations of traditional AUC evaluations by averaging per-domain AUCs and incorporating prediction polarization via Wasserstein Distance. It better reflects real-world performance under domain shifts and provides interpretable insights into detector degradation.
Dango is a 1.8B-parameter LLM designed to study Japanese-to-English second language acquisition. It uses a filtering method to minimize English contamination in monolingual pretraining, preserving realistic L1 exposure. Fine-tuned on LLM-generated lessons, Dango produces human-like L2 outputs, outperforming unfiltered and standard multilingual models.
This paper introduces StruBI, an algorithm that identifies hidden confounding and selection biases by analyzing causal mechanism shifts across environments. It formalizes a mutual information-based criterion to detect structural biases and demonstrates superior performance in recovering biased variables on synthetic and real-world data.
A meta-feedback semi-supervised learning framework enables scaling of automatic emergency braking using massive unlabeled fleet data. The stabilized approach reduces pseudo-label errors and suppresses risk hallucinations, achieving a 100:1 positive-to-false activation ratio and 35% more accident-free driving mileage compared to a rule-only baseline in real-world deployment.
TransitNet, a compact attention-augmented deep learning framework, achieves 95.2% accuracy in low-SNR transit blind searches, outperforming TLS and BLS in ROC-AUC and PR-AP values. It recovers 93.0% of injected Earth- and sub-Earth-size transits, with 97.4% of injected transits fully covered by estimated transit windows, and successfully recovers all 34 confirmed Kepler planets with a mean midpoint error of 1.24 hours.
A new framework enables zero-shot active feature acquisition by leveraging LLMs to elicit only discriminative statistics like unary deviations and pairwise co-variations. Using maximum-entropy closure, it resolves ambiguity in feature selection and outperforms LLMs itself, especially on challenging IBD patient cases where diagnostic uncertainty is high.
DIPHINE is the first neural estimator that uses score-based diffusion models to jointly estimate all mutual information terms required by Integrated Information Decomposition ($Φ$ID) from a single amortized network. It recovers the sixteen non-overlapping information atoms via Möbius inversion and provides a theoretical analysis showing synergy-to-synergy estimation is the most challenging, with accurate results on synthetic benchmarks and real-world physiological data.
Sumi is a 7B-parameter uniform diffusion language model pretrained from scratch on 1.5T tokens. It competes with autoregressive models on knowledge, reasoning, and coding tasks but underperforms on commonsense benchmarks, likely due to its education-heavy data mixture. The model weights, checkpoints, and full training recipe are publicly released.
This paper analyzes discontinuities in Sparse Mixture-of-Experts models, classifying them by order and showing that lower-order discontinuities dominate in volume. It proves that random input paths almost surely first hit an order-1 discontinuity with finite-time probability bounds and derives occupation-time bounds for each order. A simple smoothing mechanism is proposed that enhances model continuity and performance with minimal computational overhead.
A new framework uses positive-unlabeled learning and Partial Optimal Transport to audit LLM evaluation biases. It aligns human-verified positive outputs with unlabelled model responses in embedding space, identifying consistent human preferences and correcting verbosity bias without retraining. Experiments show improved human alignment, robustness to presentation biases, and interpretable confidence estimates.
Giskard enables confidential and Byzantine-robust decentralized machine learning aggregation by organizing parties into tree-based committees of size O(log n). It uses BGW-style MPC and a committee-adapted binary search to compute an approximate median, reducing per-party communication complexity asymptotically while maintaining model utility under up to n/4 Byzantine parties.
OrthoReg introduces orthogonal regularization to prevent neural components from relearning symbolic structures in hybrid dynamical systems. By directly penalizing overlap between symbolic and neural parts, it enables a complementary decomposition where symbolic models capture expressible physics and neural models handle remaining dynamics. On benchmarks with partial library mismatch, OrthoReg improves symbolic recovery and out-of-distribution performance.
Morpheus is a morphology-aware neural tokenizer and word embedder for Turkish that preserves original text through lossless encoding and decoding. It achieves the lowest bits-per-character (1.425), improves morphological alignment (MorphScore macro-F1 0.61), and uses 19% less GPU memory than 64K-vocabulary subword tokenizers. Frozen Morpheus embeddings outperform BGE-M3 and BERTurk in lexical retrieval, with root-family MAP of 0.85 and ROC-AUC of 1.00.
LegalWorld models Chinese civil litigation as a causally connected chain of five stages, based on 75,309 judgments. It includes reusable infrastructure to maintain consistency across stages and enables LongJud-Bench to evaluate agent performance across all phases, revealing significant capability gaps between models in different legal tasks.
A meta-feedback semi-supervised learning framework enables scaling of automatic emergency braking using massive unlabeled fleet data. The stabilized approach reduces pseudo-label errors through noise-aware decoupling and kinematics-gated pseudo-labeling, improving safety with a 100:1 positive-to-false activation ratio and 35% more accident-free driving mileage compared to rule-based systems.