ViCLSR adapts supervised contrastive learning (SimCSE-style) to Vietnamese NLU by converting NLI entailment and contradiction pairs into positive and negative training signals. Built on XLM-R Large (550M), the framework improves sentence embeddings for low-resource Vietnamese, reporting gains of +6.97% F1 over PhoBERT on ViNLI and state-of-the-art results across five downstream tasks including fact-checking and machine reading comprehension.

This paper investigates whether neural text-to-speech systems capture consonant-induced F0 perturbation—fine-grained phonetic effects where voiceless obstruents raise and voiced obstruents lower fundamental frequency relative to sonorants. The authors propose a segmental-level prosodic probing framework comparing Tacotron 2 and FastSpeech 2 against natural speech, stratifying by lexical frequency to test memorization versus abstraction. This matters because TTS evaluation often misses sub-phonemic articulatory detail that distinguishes human-like phonetic competence from surface pattern matching.

This paper evaluates three inference-time strategies—self-consistency with temperature/top-p sampling, dual-model cross-verification, and iterative self-reflection—to improve multi-step reasoning in LLMs without parameter updates. The core premise is that aggregating diverse reasoning traces or validating across models yields more reliable outputs than single-pass decoding. The work addresses a practical need for deployment scenarios where retraining is infeasible, though the experimental scope is limited by unclear model specifications and dataset choices.

This paper tackles multimodal hate speech detection where hateful intent emerges from complex interactions between text and images—what the authors call "more than the sum of its parts." The core innovation is the Stratified Multimodal Interaction (SMI) paradigm, which categorizes eight distinct cross-modal interaction patterns into three difficulty levels (Easy, Normal, Hard), coupled with the ARCADE framework that simulates an asymmetric courtroom debate between Prosecutor, Defender, and Judge agents to decipher subtle intent shifts. This matters because current detection systems fail when hateful content is constructed implicitly through benign-seeming modalities that only become toxic in combination.

This paper tackles logical context poisoning—the degradation of LLM responses when flat, linear conversation structures force topically distinct threads to accumulate in a single unbounded context window. The core idea is the Conversation Tree Architecture (CTA), which models conversations as a directed rooted tree $\mathcal{T}=(V,E,r,W)$ where each node $v \in V$ maintains an isolated local context window $w_v$. Structured flow operations—downstream passing $\phi_{\downarrow}$, upstream merging $\psi_{\uparrow}$, and volatile nodes—govern how context moves between branches. This matters because current interfaces offer no middle ground between discarding context (new chat) and accumulating noise (linear threads).

Extracting hypotheses and their supporting statistical evidence from full-text scientific articles is challenging due to document length and the distribution of scientific arguments across sections. This paper proposes a two-stage retrieve-and-extract pipeline that first links an abstract finding to its corresponding hypothesis, then extracts the statistical evidence supporting that hypothesis. Through controlled ablations varying context quantity ($k \in \{5, 10, 20\}$), retrieval quality (standard RAG, reranking, fine-tuned retriever), and oracle paragraph settings, the authors demonstrate that hypothesis extraction is primarily bounded by retrieval quality, while evidence extraction faces persistent extractor limitations even with perfect paragraph selection.

This paper formalizes the transformer context window as an I/O page and proves that tool-augmented agents with indexed external memory achieve exponential retrieval cost savings over sequential scanning: $\mathcal{O}(\log_b N)$ versus $\Omega(N)$ page reads. The authors validate these predictions experimentally across three content types and identify "parametric memory competition" as a failure mode where models bypass retrieval protocols for familiar content.

LongCat-Flash-Prover is a 560B-parameter MoE open-source model targeting native formal reasoning in Lean4. The core innovation is decomposing formal theorem proving into three agentic capabilities—auto-formalization, sketching, and proving—trained via a Hybrid-Experts Iteration Framework and a novel RL algorithm called HisPO. The work claims state-of-the-art results on MiniF2F-Test (97.1%), ProverBench (70.8%), and PutnamBench (41.5%) with remarkably low inference budgets compared to prior open-source provers.

This paper addresses selection bias (position and label bias) in large language models during discrete-choice tasks like multiple-choice questions and pairwise evaluation. The authors propose Permutation-Aware GRPO (PA-GRPO), which extends Group Relative Policy Optimization by treating different permutations of the same question as a single training group rather than independent instances. The method enforces semantic consistency across permutations through two mechanisms: a cross-permutation advantage that computes rewards relative to the group mean, and a consistency-aware reward that penalizes disagreement across permutations. Experiments across seven benchmarks and three models (Llama-3.1-8B, Qwen3-8B, Qwen3-32B) demonstrate that PA-GRPO reduces selection bias while maintaining accuracy.

This paper investigates why compressing different weight matrices in transformers leads to wildly different outcomes—from negligible impact to 20,000× perplexity increases. The authors map this structural sensitivity across five architectures, revealing that early-layer MLP up-projections are catastrophically fragile while value projections are nearly free to compress. Using Lyapunov stability theory, they explain how residual connections contract errors, and they provide machine-checked formal bounds in Lean 4 to guarantee per-matrix approximation quality.

NVIDIA introduces Nemotron 3, a family of open language models (Nano, Super, Ultra) built on a hybrid Mamba-Transformer MoE architecture. The core innovation is using selective attention layers combined with Mamba-2 state space layers to achieve high throughput while maintaining accuracy. Key technical contributions include LatentMoE (dimensionality-reduced expert routing), NVFP4 training for efficiency, and multi-environment RL post-training. The paper positions these models as optimized for agentic AI with up to 1M token contexts and granular inference-time reasoning budget control.