Uncategorized

arXiv:2601.04411v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards (RLVR) is a simple but powerful paradigm for training LLMs: sample a completion, verify it, and update. In practice, however, the verifier is almost never clean–unit tests probe only limited corner cases; human and synthetic labels are imperfect; and LLM judges (e.g., RLAIF) are noisy and can be exploited–and this problem worsens on harder domains (especially coding) where tests are sparse and increasingly model-generated. We ask a pragmatic question: does the verification noise merely slow down the learning (rate), or can it flip the outcome (fate)? To address this, we develop an analytically tractable multi-armed bandit view of RLVR dynamics, instantiated with GRPO and validated in controlled experiments. Modeling false positives and false negatives and grouping completions into recurring reasoning modes yields a replicator-style (natural-selection) flow on the probability simplex. The dynamics decouples into within-correct-mode competition and a one-dimensional evolution for the mass on incorrect modes, whose drift is determined solely by Youden’s index J=TPR-FPR. This yields a sharp phase transition: when J>0, the incorrect mass is driven toward extinction (learning); when J=0, the process is neutral; and when J0, noise primarily rescales convergence time (“rate, not fate”). Experiments on verifiable programming tasks under synthetic noise reproduce the predicted J=0 boundary. Beyond noise, the framework offers a general lens for analyzing RLVR stability, convergence, and algorithmic interventions.

arXiv:2512.21002v2 Announce Type: replace Abstract: Distilling the capabilities from a large reasoning model (LRM) to a smaller student model often involves training on substantial amounts of reasoning data. However, knowledge distillation (KD) over lengthy sequences with prompt (P), chain-of-thought (CoT), and answer (A) sections makes the process computationally expensive. In this work, we investigate how the allocation of supervision across different sections (P, CoT, A) affects student performance. Our analysis shows that selective KD over only the CoT tokens can be effective when the prompt and answer information is encompassed by it. Building on this insight, we establish a truncation protocol to quantify computation-quality tradeoffs as a function of sequence length. We observe that beyond a specific length, longer training sequences provide marginal returns for downstream performance but require substantially higher memory and FLOPs. To this end, training on only the first $50%$ of tokens of every training sequence can retain, on average, $approx91%$ of full-sequence performance on math benchmarks while reducing training time, memory usage, and FLOPs by about $50%$ each. Codes are available at https://github.com/weiruichen01/distilling-the-essence.

arXiv:2601.03615v1 Announce Type: new Abstract: Audio Language Models (ALMs) offer a promising shift towards explainable audio deepfake detections (ADDs), moving beyond textit{black-box} classifiers by providing some level of transparency into their predictions via reasoning traces. This necessitates a new class of model robustness analysis: robustness of the predictive reasoning under adversarial attacks, which goes beyond existing paradigm that mainly focuses on the shifts of the final predictions (e.g., fake v.s. real). To analyze such reasoning shifts, we introduce a forensic auditing framework to evaluate the robustness of ALMs’ reasoning under adversarial attacks in three inter-connected dimensions: acoustic perception, cognitive coherence, and cognitive dissonance. Our systematic analysis reveals that explicit reasoning does not universally enhance robustness. Instead, we observe a bifurcation: for models exhibiting robust acoustic perception, reasoning acts as a defensive textit{“shield”}, protecting them from adversarial attacks. However, for others, it imposes a performance textit{“tax”}, particularly under linguistic attacks which reduce cognitive coherence and increase attack success rate. Crucially, even when classification fails, high cognitive dissonance can serve as a textit{silent alarm}, flagging potential manipulation. Overall, this work provides a critical evaluation of the role of reasoning in forensic audio deepfake analysis and its vulnerabilities.

arXiv:2601.03682v1 Announce Type: new Abstract: Recent studies reveal that large language models (LLMs) exhibit limited logical reasoning abilities in mathematical problem-solving, instead often relying on pattern-matching and memorization. We systematically analyze this limitation, focusing on logical relationship understanding, which is a core capability underlying genuine logical reasoning, and reveal that errors related to this capability account for over 90% of incorrect predictions, with Chain-of-Thought Supervised Fine-Tuning (CoT-SFT) failing to substantially reduce these errors. To address this bottleneck, we propose First-Step Logical Reasoning (FSLR), a lightweight training framework targeting logical relationship understanding. Our key insight is that the first planning step-identifying which variables to use and which operation to apply-encourages the model to derive logical relationships directly from the problem statement. By training models on this isolated step, FSLR provides explicit supervision for logical relationship understanding, unlike CoT-SFT which implicitly embeds such relationships within complete solution trajectories. Extensive experiments across multiple models and datasets demonstrate that FSLR consistently outperforms CoT-SFT under both in-distribution and out-of-distribution settings, with average improvements of 3.2% and 4.6%, respectively. Moreover, FSLR achieves 4-6x faster training and reduces training token consumption by over 80%.

arXiv:2601.03605v1 Announce Type: new Abstract: Despite the significant advancements of Large Language Models (LLMs), their factuality remains a critical challenge, fueling growing interest in factuality verification. Existing research on factuality verification primarily conducts binary judgments (e.g., correct or incorrect), which fails to distinguish varying degrees of error severity. This limits its utility for applications such as fine-grained evaluation and preference optimization. To bridge this gap, we propose the Agentic Discriminative Verifier (DiVA), a hybrid framework that synergizes the agentic search capabilities of generative models with the precise scoring aptitude of discriminative models. We also construct a new benchmark, FGVeriBench, as a robust testbed for fine-grained factuality verification. Experimental results on FGVeriBench demonstrate that our DiVA significantly outperforms existing methods on factuality verification for both general and multi-hop questions.

arXiv:2512.17394v2 Announce Type: replace Abstract: Theory of Mind (ToM) – the ability to attribute beliefs and intents to others – is fundamental for social intelligence, yet Vision-Language Model (VLM) evaluations remain largely Western-centric. In this work, we introduce CulturalToM-VQA, a benchmark of 5,095 visually situated ToM probes across diverse cultural contexts, rituals, and social norms. Constructed through a frontier proprietary MLLM, human-verified pipeline, the dataset spans a taxonomy of six ToM tasks and four complexity levels. We benchmark 10 VLMs (2023-2025) and observe a significant performance leap: while earlier models struggle, frontier models achieve high accuracy (>93%). However, significant limitations persist: models struggle with false belief reasoning (19-83% accuracy) and show high regional variance (20-30% gaps). Crucially, we find that SOTA models exhibit social desirability bias – systematically favoring semantically positive answer choices over negative ones. Ablation experiments reveal that some frontier models rely heavily on parametric social priors, frequently defaulting to safety-aligned predictions. Furthermore, while Chain-of-Thought prompting aids older models, it yields minimal gains for newer ones. Overall, our work provides a testbed for cross-cultural social reasoning, underscoring that despite architectural gains, achieving robust, visually grounded understanding remains an open challenge.

arXiv:2512.22431v2 Announce Type: replace-cross Abstract: The proliferation of Large Language Models (LLMs) has catalyzed a shift towards autonomous agents capable of complex reasoning and tool use. However, current agent architectures are frequently constructed using imperative, ad hoc patterns. This results in brittle systems plagued by difficulties in state management, error handling, and concurrency. This paper introduces Monadic Context Engineering (MCE), a novel architectural paradigm leveraging the algebraic structures of Functors, Applicative Functors, and Monads to provide a formal foundation for agent design. MCE treats agent workflows as computational contexts where cross-cutting concerns, such as state propagation, short-circuiting error handling, and asynchronous execution, are managed intrinsically by the algebraic properties of the abstraction. We demonstrate how Monads enable robust sequential composition, how Applicatives provide a principled structure for parallel execution, and crucially, how Monad Transformers allow for the systematic composition of these capabilities. This layered approach enables developers to construct complex, resilient, and efficient AI agents from simple, independently verifiable components. We further extend this framework to describe Meta-Agents, which leverage MCE for generative orchestration, dynamically creating and managing sub-agent workflows through metaprogramming.

arXiv:2512.07515v2 Announce Type: replace Abstract: Detecting hallucinations in Retrieval-Augmented Generation remains a challenge. Prior approaches attribute hallucinations to a binary conflict between internal knowledge stored in FFNs and the retrieved context. However, this perspective is incomplete, failing to account for the impact of other components of the LLM, such as the user query, previously generated tokens, the self token, and the final LayerNorm adjustment. To comprehensively capture the impact of these components on hallucination detection, we propose TPA which mathematically attributes each token’s probability to seven distinct sources: Query, RAG Context, Past Token, Self Token, FFN, Final LayerNorm, and Initial Embedding. This attribution quantifies how each source contributes to the generation of the next token. Specifically, we aggregate these attribution scores by Part-of-Speech (POS) tags to quantify the contribution of each model component to the generation of specific linguistic categories within a response. By leveraging these patterns, such as detecting anomalies where Nouns rely heavily on LayerNorm, TPA effectively identifies hallucinated responses. Extensive experiments show that TPA achieves state-of-the-art performance.

arXiv:2512.17843v2 Announce Type: replace Abstract: While academic research typically treats Large Language Models (LLM) as generic text generators, they are distinct commercial products with unique interfaces and capabilities that fundamentally shape user behavior. Current datasets obscure this reality by collecting text-only data through uniform interfaces that fail to capture authentic chatbot usage. To address this limitation, we present ShareChat, a large-scale corpus of 142,808 conversations (660,293 turns) sourced directly from publicly shared URLs on ChatGPT, Perplexity, Grok, Gemini, and Claude. ShareChat distinguishes itself by preserving native platform affordances, such as citations and thinking traces, across a diverse collection covering 101 languages and the period from April 2023 to October 2025. Furthermore, ShareChat offers substantially longer context windows and greater interaction depth than prior datasets. To illustrate the dataset’s breadth, we present three case studies: a completeness analysis of intent satisfaction, a citation study of model grounding, and a temporal analysis of engagement rhythms. This work provides the community with a vital and timely resource for understanding authentic user-LLM chatbot interactions in the wild. The dataset will be publicly available.

arXiv:2601.02609v1 Announce Type: cross Abstract: Large language model fine-tuning is bottlenecked by memory: a 7B parameter model requires 84GB–14GB for weights, 14GB for gradients, and 56GB for FP32 optimizer states–exceeding even A100-40GB capacity. We present Chronicals, an open-source training framework achieving 3.51x speedup over Unsloth through four synergistic optimizations: (1) fused Triton kernels eliminating 75% of memory traffic via RMSNorm (7x), SwiGLU (5x), and QK-RoPE (2.3x) fusion; (2) Cut Cross-Entropy reducing logit memory from 5GB to 135MB through online softmax computation; (3) LoRA+ with theoretically-derived 16x differential learning rates between adapter matrices; and (4) Best-Fit Decreasing sequence packing recovering 60-75% of compute wasted on padding. On Qwen2.5-0.5B with A100-40GB, Chronicals achieves 41,184 tokens/second for full fine-tuning versus Unsloth’s 11,736 tokens/second (3.51x). For LoRA at rank 32, we reach 11,699 tokens/second versus Unsloth MAX’s 2,857 tokens/second (4.10x). Critically, we discovered that Unsloth’s reported 46,000 tokens/second benchmark exhibited zero gradient norms–the model was not training. We provide complete mathematical foundations: online softmax correctness proofs, FlashAttention IO complexity bounds O(N^2 d^2 M^{-1}), LoRA+ learning rate derivations from gradient magnitude analysis, and bin-packing approximation guarantees. All implementations, benchmarks, and proofs are available at https://github.com/Ajwebdevs/Chronicals with pip installation via https://pypi.org/project/chronicals/.