Uncategorized

arXiv:2602.22897v1 Announce Type: cross Abstract: Human intelligence naturally intertwines omni-modal perception — spanning vision, audio, and language — with complex reasoning and tool usage to interact with the world. However, current multi-modal LLMs are primarily confined to bi-modal interactions (e.g., vision-language), lacking the unified cognitive capabilities required for general AI assistants. To bridge this gap, we introduce OmniGAIA, a comprehensive benchmark designed to evaluate omni-modal agents on tasks necessitating deep reasoning and multi-turn tool execution across video, audio, and image modalities. Constructed via a novel omni-modal event graph approach, OmniGAIA synthesizes complex, multi-hop queries derived from real-world data that require cross-modal reasoning and external tool integration. Furthermore, we propose OmniAtlas, a native omni-modal foundation agent under tool-integrated reasoning paradigm with active omni-modal perception. Trained on trajectories synthesized via a hindsight-guided tree exploration strategy and OmniDPO for fine-grained error correction, OmniAtlas effectively enhances the tool-use capabilities of existing open-source models. This work marks a step towards next-generation native omni-modal AI assistants for real-world scenarios.

arXiv:2602.21262v2 Announce Type: replace Abstract: With increasing integration of Large Language Models (LLMs) into areas of high-stakes human decision-making, it is important to understand the risks they introduce as advisors. To be useful advisors, LLMs must sift through large amounts of content, written with both benevolent and malicious intent, and then use this information to convince a user to take a specific action. This involves two social capacities: vigilance (the ability to determine which information to use, and which to discard) and persuasion (synthesizing the available evidence to make a convincing argument). While existing work has investigated these capacities in isolation, there has been little prior investigation of how these capacities may be linked. Here, we use a simple multi-turn puzzle-solving game, Sokoban, to study LLMs’ abilities to persuade and be rationally vigilant towards other LLM agents. We find that puzzle-solving performance, persuasive capability, and vigilance are dissociable capacities in LLMs. Performing well on the game does not automatically mean a model can detect when it is being misled, even if the possibility of deception is explicitly mentioned. However, LLMs do consistently modulate their token use, using fewer tokens to reason when advice is benevolent and more when it is malicious, even if they are still persuaded to take actions leading them to failure. To our knowledge, our work presents the first investigation of the relationship between persuasion, vigilance, and task performance in LLMs, and suggests that monitoring all three independently will be critical for future work in AI safety.

arXiv:2602.22698v1 Announce Type: new Abstract: Leveraging Large Language Models (LLMs) for Knowledge Graph Completion (KGC) is promising but hindered by a fundamental granularity mismatch. LLMs operate on fragmented token sequences, whereas entities are the fundamental units in knowledge graphs (KGs) scenarios. Existing approaches typically constrain predictions to limited candidate sets or align entities with the LLM’s vocabulary by pooling multiple tokens or decomposing entities into fixed-length token sequences, which fail to capture both the semantic meaning of the text and the structural integrity of the graph. To address this, we propose KGT, a novel framework that uses dedicated entity tokens to enable efficient, full-space prediction. Specifically, we first introduce specialized tokenization to construct feature representations at the level of dedicated entity tokens. We then fuse pre-trained structural and textual features into these unified embeddings via a relation-guided gating mechanism, avoiding training from scratch. Finally, we implement decoupled prediction by leveraging independent heads to separate and combine semantic and structural reasoning. Experimental results show that KGT consistently outperforms state-of-the-art methods across multiple benchmarks.

arXiv:2602.22658v1 Announce Type: cross Abstract: Deepfake speech utterances can be forged by replacing one or more words in a bona fide utterance with semantically different words synthesized by speech generative models. While a dedicated synthetic word detector could be developed, we investigate a cost-effective method that fine-tunes a pre-trained Whisper model to detect synthetic words while transcribing the input utterance via next-token prediction. We further investigate using partially vocoded utterances as the fine-tuning data, thereby reducing the cost of data collection. Our experiments demonstrate that, on in-domain test data, the fine-tuned Whisper yields low synthetic-word detection error rates and transcription error rates. On out-of-domain test data with synthetic words produced by unseen speech generative models, the fine-tuned Whisper remains on par with a dedicated ResNet-based detection model; however, the overall performance degradation calls for strategies to improve its generalization capability.

arXiv:2602.23070v1 Announce Type: cross Abstract: Although Automatic Speech Recognition (ASR) in Bengali has seen significant progress, processing long-duration audio and performing robust speaker diarization remain critical research gaps. To address the severe scarcity of joint ASR and diarization resources for this language, we introduce Lipi-Ghor-882, a comprehensive 882-hour multi-speaker Bengali dataset. In this paper, detailing our submission to the DL Sprint 4.0 competition, we systematically evaluate various architectures and approaches for long-form Bengali speech. For ASR, we demonstrate that raw data scaling is ineffective; instead, targeted fine-tuning utilizing perfectly aligned annotations paired with synthetic acoustic degradation (noise and reverberation) emerges as the singular most effective approach. Conversely, for speaker diarization, we observed that global open-source state-of-the-art models (such as Diarizen) performed surprisingly poorly on this complex dataset. Extensive model retraining yielded negligible improvements; instead, strategic, heuristic post-processing of baseline model outputs proved to be the primary driver for increasing accuracy. Ultimately, this work outlines a highly optimized dual pipeline achieving a $sim$0.019 Real-Time Factor (RTF), establishing a practical, empirically backed benchmark for low-resource, long-form speech processing.

arXiv:2602.22225v1 Announce Type: cross Abstract: Retrieval-augmented generation (RAG) has strong potential for producing accurate and factual outputs by combining language models (LMs) with evidence retrieved from large text corpora. However, current pipelines are limited by static chunking and flat retrieval: documents are split into short, predetermined, fixed-size chunks, embeddings are retrieved uniformly, and generation relies on whatever chunks are returned. This design brings challenges, as retrieval quality is highly sensitive to chunk size, often introduces noise from irrelevant or misleading chunks, and scales poorly to large corpora. We present SmartChunk retrieval, a query-adaptive framework for efficient and robust long-document question answering (QA). SmartChunk uses (i) a planner that predicts the optimal chunk abstraction level for each query, and (ii) a lightweight compression module that produces high-level chunk embeddings without repeated summarization. By adapting retrieval granularity on the fly, SmartChunk balances accuracy with efficiency and avoids the drawbacks of fixed strategies. Notably, our planner can reason about chunk abstractions through a novel reinforcement learning scheme, STITCH, which boosts accuracy and generalization. To reflect real-world applications, where users face diverse document types and query styles, we evaluate SmartChunk on five QA benchmarks plus one out-of-domain dataset. Across these evaluations, SmartChunk outperforms state-of-the-art RAG baselines, while reducing cost. Further analysis demonstrates strong scalability with larger corpora and consistent gains on out-of-domain datasets, highlighting its effectiveness as a general framework for adaptive retrieval.

arXiv:2602.22723v1 Announce Type: new Abstract: There is growing recognition that many NLP tasks lack a single ground truth, as human judgments reflect diverse perspectives. To capture this variation, models have been developed to predict full annotation distributions rather than majority labels, while perspectivist models aim to reproduce the interpretations of individual annotators. In this work, we compare these approaches on Implicit Discourse Relation Recognition (IDRR), a highly ambiguous task where disagreement often arises from cognitive complexity rather than ideological bias. Our experiments show that existing annotator-specific models perform poorly in IDRR unless ambiguity is reduced, whereas models trained on label distributions yield more stable predictions. Further analysis indicates that frequent cognitively demanding cases drive inconsistency in human interpretation, posing challenges for perspectivist modeling in IDRR.

arXiv:2602.22647v1 Announce Type: cross Abstract: Generative retrieval has emerged as a powerful paradigm for LLM-based recommendation. However, industrial recommender systems often benefit from restricting the output space to a constrained subset of items based on business logic (e.g. enforcing content freshness or product category), which standard autoregressive decoding cannot natively support. Moreover, existing constrained decoding methods that make use of prefix trees (Tries) incur severe latency penalties on hardware accelerators (TPUs/GPUs). In this work, we introduce STATIC (Sparse Transition Matrix-Accelerated Trie Index for Constrained Decoding), an efficient and scalable constrained decoding technique designed specifically for high-throughput LLM-based generative retrieval on TPUs/GPUs. By flattening the prefix tree into a static Compressed Sparse Row (CSR) matrix, we transform irregular tree traversals into fully vectorized sparse matrix operations, unlocking massive efficiency gains on hardware accelerators. We deploy STATIC on a large-scale industrial video recommendation platform serving billions of users. STATIC produces significant product metric impact with minimal latency overhead (0.033 ms per step and 0.25% of inference time), achieving a 948x speedup over a CPU trie implementation and a 47-1033x speedup over a hardware-accelerated binary-search baseline. Furthermore, the runtime overhead of STATIC remains extremely low across a wide range of practical configurations. To the best of our knowledge, STATIC enables the first production-scale deployment of strictly constrained generative retrieval. In addition, evaluation on academic benchmarks demonstrates that STATIC can considerably improve cold-start performance for generative retrieval. Our code is available at https://github.com/youtube/static-constraint-decoding.

Burrowed in the alleys of Hongik-dong, a hushed residential neighborhood in eastern Seoul, is a faded stone-tiled building stamped “Korea Baduk Association,” the governing body for professional Go. The game is an ancient one, with sacred stature in South Korea.  But inside the building, rooms once filled with the soft clatter of hands dipping into wooden bowls of stones now echo with mouse clicks. Players hunch over their monitors and replay their matches in an AI program. Others huddle around a Go board and debate the best next move, while coaches report how their choices stack up against the AI’s. Some sit in silence, watching AI programs play against each other.  Ten years ago AlphaGo, Google DeepMind’s AI program, stunned the world by defeating the South Korean Go player Lee Sedol. And in the years since, AI has upended the game. It’s overturned centuries-old principles about the best moves and introduced entirely new ones. Players now train to replicate AI’s moves as closely as they can rather than inventing their own, even when the machine’s thinking remains mysterious to them. Today, it is essentially impossible to compete professionally without using AI. Some say the technology has drained the game of its creativity, while others think there is still room for human invention. Meanwhile, AI is democratizing access to training, and more female players are climbing the ranks as a result.  For Shin Jin-seo, the top-ranked Go player in the world, AI is an invaluable training partner. Every morning, he sits at his computer and opens a program called KataGo. Nicknamed “Shintelligence” for how closely his moves mimic AI’s, he traces the glowing “blue spot” that represents the program’s suggestion for the best next move, rearranging the stones on the digital grid to try to understand the machine’s thinking. “I constantly think about why AI chose a move,” he says. When training for a match, Shin spends most of his waking hours poring over KataGo. “It’s almost like an ascetic practice,” he says. According to a study in 2022 by the Korean Baduk League, Shin’s moves match AI’s 37.5% of the time, well above the 28.5% average the study found among all players. “My game has changed a lot,” says Shin, “because I have to follow the directions suggested by AI to some extent.” The Korea Baduk Association says it has reached out to Google DeepMind in the hopes of arranging a match between Shin and AlphaGo, to commemorate the 10th anniversary of its victory over Lee. A spokesperson for Google DeepMind said the company could not provide information at this time. But if a new match does happen, Shin, who has trained on more advanced AI programs, is optimistic that he’d win. “AlphaGo still had some flaws then, so I think I could beat it if I target those weaknesses,” he says. AI rewrites the Go playbook Go is an abstract strategy board game invented in China more than 2,500 years ago. Two players take turns placing black and white stones on a 19×19 grid, aiming to conquer territory by surrounding their opponent’s stones. It’s a game of striking mathematical complexity. The number of possible board configurations—roughly 10170—dwarfs the number of atoms in the universe. If chess is a battle, Go is a war. You suffocate your enemy in one corner while fending off an invasion in another. To train AI to play Go, a vast trove of human Go moves are fed into a neural network, a computing system that mimics the web of neurons in the human brain. AlphaGo, which was later christened AlphaGo Lee after its victory over Lee Sedol, was trained on 30 million Go moves and refined by playing millions of games against itself. In 2017, its successor, AlphaGo Zero, picked up Go from scratch. Without studying any human games, it learned by playing against itself, with moves based only on the rules of the game. The blank-slate approach proved more powerful, unconstrained by the limits of human knowledge. After three days of training, it beat AlphaGo Lee 100 games to zero.  Google DeepMind retired AlphaGo that same year. But then a wave of open-source models inspired by AlphaGo Zero emerged. Today, KataGo is the program most widely used by professional Go players in South Korea. It’s faster and sharper than AlphaGo. It’s learned to predict not just who will win, but also who owns each point on the board at any given moment. While AlphaGo Zero pieced together its understanding of the board by looking at small sections, KataGo learned to read the whole board, developing better judgment for long-term strategies. Instead of just learning how to win, it learned to maximize its score. The software has reshaped how people play. For hundreds of years, professional Go players have navigated the game’s astronomical complexity by developing heuristics that replaced brute calculation. Elegant opening strategies imposed abstract order on the empty grid. Invading corners early was a bad bargain. Each generation of Go players added new principles to the canon.  But “AI has changed everything,” says Park Jeong-sang, a South Korean Go commentator. “Fundamental moves that were once considered common sense aren’t played at all today, and techniques that didn’t exist before have become popular.”  The starkest shift has been in opening moves. Go starts on a blank grid, and the first 50 moves were canvases for abstract thinking and creativity, where players etched their personalities and philosophies. Lee Sedol fashioned provocative moves that invited chaos. Ke Jie, a Chinese player who was defeated by AlphaGo Master in 2017, dazzled with agile, imaginative moves. Now, players memorize the same strain of efficient, calculated opening moves suggested by AI. The crux of the game has shifted to the middle moves, where raw calculation matters more than creativity. Training with AI has led to a homogenization of playing styles. Ke Jie has lamented the strain of watching the same opening moves recycled endlessly. “I feel the exact same way as the fans watching. It’s very tiring

arXiv:2506.09886v2 Announce Type: replace Abstract: Detecting hallucinations in large language models (LLMs) is critical for their safety in many applications. Without proper detection, these systems often provide harmful, unreliable answers. In recent years, LLMs have been actively used in retrieval-augmented generation (RAG) settings. However, hallucinations remain even in this setting, and while numerous hallucination detection methods have been proposed, most approaches are not specifically designed for RAG systems. To overcome this limitation, we introduce a hallucination detection method based on estimating the distances between the distributions of prompt token embeddings and language model response token embeddings. The method examines the geometric structure of token hidden states to reliably extract a signal of factuality in text, while remaining friendly to long sequences. Extensive experiments demonstrate that our method achieves state-of-the-art or competitive performance. It also has transferability from solving the NLI task to the hallucination detection task, making it a fully unsupervised and efficient method with a competitive performance on the final task.