Abstract

How should we evaluate generation systems that combine autoregressive (AR) and diffusion decoding? We study this question through Speculative Refinement (SpecRef), a training-free hybrid method that warm-starts a masked diffusion language model from an AR draft using entropy-guided selective masking. Evaluating SpecRef across six benchmarks (HumanEval, MBPP, GSM8K, BBH, ARC-Challenge, HellaSwag) with three distinct evaluation protocols (execution-based pass@1, exact-match, log-likelihood scoring), we surface several findings relevant beyond our specific system: (1) code benchmarks conflate structural discovery} with logical correctness: providing a syntactic scaffold lifts accuracy from near zero to over 20% without changing the model, indicating that much of the baseline failure is structural; (2) a refinement tension phenomenon where multi-stage correction degrades already-correct tokens, exposing benchmark saturation ceilings invisible to single-model evaluation; (3) log-likelihood and generative evaluation produce different model rankings for the same model pair, suggesting they measure different capabilities; (4) standard Python post-processing silently breaks code evaluation for non-AR generators. These observations apply to any multi-stage or non-autoregressive generation pipeline and point toward more diagnostic evaluation practices.

Abstract

Raz (2009) proved that multilinear formulas computing the determinant of a generic $n times n$ matrix require size $n^{Omega(log n)}$. A fundamental question in understanding this lower bound is identifying which structural properties of the determinant drive this hardness. Is it the quadratic number of variables? The dense connectivity? Specific algebraic symmetries?

We establish that density is not essential. We prove the existence of $n times n$ symbolic matrices with only $Theta(n log^6 n)$ nonzero entries - reducing the variable count by a factor of $n/{log^6 n}$ - such that any multilinear formula computing their determinants still requires size $n^{Omega(log n)}$. Our construction uses rectangle sampling from the complete bipartite graph to generate sparse matrices that simultaneously maintain perfect matchings (ensuring nonzero determinant) while exhibiting diagonal imbalance under random vertex permutations - a geometric property we identify as the key driver of factor imbalance in Raz's framework.

This demonstrates that Raz's partial derivatives method is remarkably robust to sparsification, and suggests that the fundamental source of multilinear hardness for determinant lies in expansion-like combinatorial structure rather than density. Our techniques combine concentration inequalities for dependent random variables with insights from random graph theory.

Abstract

Large Language Models often achieve strong performance by generating long intermediate chain-of-thought reasoning. However, it remains unclear when a model's final answer is actually determined during generation. If the answer is already fixed at an intermediate stage, subsequent reasoning tokens may constitute post-decision explanation, increasing inference cost and latency without improving correctness. We study the evolution of predicted answers over reasoning steps using forced answer completion, which elicits the model's intermediate predictions at partial reasoning prefixes. Focusing on Qwen3-4B and averaging results across all datasets considered, we find that predicted answers change in only 32% of queries. Moreover, once the final answer switch occurs, the model generates an average of 760 additional reasoning tokens per query, accounting for a substantial fraction of the total reasoning budget. Motivated by these findings, we investigate early stopping strategies that halt generation once the answer has stabilized. We show that simple heuristics, including probe-based stopping, can reduce reasoning token usage by 500 tokens per query while incurring only a 2% drop in accuracy. Together, our results indicate that a large portion of chain-of-thought generation is redundant and can be reduced with minimal impact on performance.

Abstract

Large language models (LLMs) exhibit failures on elementary symbolic tasks such as character counting in a word, despite excelling on complex benchmarks. Although this limitation has been noted, the internal reasons remain unclear. We use character counting (e.g., "How many p's are in apple?") as a minimal, controlled probe that isolates token-level reasoning from higher-level confounds. Using this setting, we uncover a consistent phenomenon across modern architectures, including LLaMA, Qwen, and Gemma: models often compute the correct answer internally yet fail to express it at the output layer.
Through mechanistic analysis combining probing classifiers, activation patching, logit lens analysis, and attention head tracing, we show that character-level information is encoded in early and mid-layer representations. However, this information is attenuated by a small set of components in later layers, especially the penultimate and final layer MLP. We identify these components as negative circuits: subnetworks that downweight correct signals in favor of higher-probability but incorrect outputs.
Our results lead to two contributions. First, we show that symbolic reasoning failures in LLMs are not due to missing representations or insufficient scale, but arise from structured interference within the model's computation graph. This explains why such errors persist and can worsen under scaling and instruction tuning. Second, we provide evidence that LLM forward passes implement a form of competitive decoding, in which correct and incorrect hypotheses coexist and are dynamically reweighted, with final outputs determined by suppression as much as by amplification.
These findings carry implications for interpretability and robustness: simple symbolic reasoning exposes weaknesses in modern LLMs, underscoring need for design strategies that ensure information is encoded and reliably used.

Abstract

We provide a comprehensive review of the FRUIT (Faithfully Reflecting Updated Information in Text) task, which formalizes the challenge of accurately updating textual information with large language models (LLMs). Our work begins with an in-depth analysis of the FRUIT dataset, revealing key structural insights. We also investigate the unsupervised capabilities of LLMs--such as zero-shot learning, chain-of-thought reasoning, self-reflection, and evidence ordering. Experimental results demonstrate that unsupervised approaches perform competitively with supervised methods in faithful text updating. Qualitative analysis shows that updates utilizing table-structured evidence outperform those based on unstructured text. We also discuss important limitations, including the need for new datasets and the risks of information leakage in this domain. These findings have significant implications for applications requiring precise document updates, such as software engineering, technical documentation, and legal document maintenanc

Abstract

Reliable recognition of motorcycle driving events is important for rider safety and intelligent mobility applications. This work addresses the limited generalisability of earlier studies by introducing a more diverse real-world time-series dataset collected using 3-axis accelerometer and gyroscope sensors from both male and female riders, recorded during day and night riding. The dataset captures wider variations in riding behaviour and sensor dynamics, making event recognition more realistic and challenging. To model this increased variability, we focus on a Time-Series Transformer (TST) architecture that can capture long-range temporal dependencies while remaining efficient for embedded deployment. The proposed approach is evaluated on edge hardware, where inference latency, memory usage, and model size are analysed. Experimental results show that the TST achieves strong and consistent class-wise performance while maintaining low computational overhead, making it well suited for real-time on-device two-wheeler event recognition.

Abstract

Automatic syllable stress detection is a key component of computer-assisted language learning (CALL) systems, as it enables precise feedback to second-language (L2) learners. Most of the existing detection methods depend on explicit phoneme or syllable-level boundaries derived from manual annotation or forced alignment. However, manual annotation is costly, and the forced alignment methods are prone to errors. To overcome these limitations, we introduce a fully boundary-independent framework for syllable stress detection based on hierarchical time-compression. Our approach employs an encoder-decoder architecture with Hierarchical compact self-Attention based Time-compression (HAT) to progressively compress the frame-level embeddings into syllable-level representations. This enables the model to capture stress cues directly from speech without any boundary supervision. We evaluate the proposed framework on the ISLE corpus of second language English learners (German and Italian), leveraging self-supervised wav2vec 2.0 embeddings. Experimental results show that our approach consistently outperforms both boundary-dependent and prior partially boundary-independent baselines. Our approach achieves a maximum accuracy of 95.07% on GER and 96.18% on ITA, with highest absolute improvements of up to 23% and 19%.

Abstract

Climate change debates unfold increasingly on social media platforms, with YouTube
serving as both a news source and a space for public discourse. While prior studies have
often examined climate discourse at a global level, less attention has been paid to how
geopolitical divides shape narratives and public responses online. This paper presents an
exploratory analysis of climate-related YouTube videos through the lens of the Global
North-South divide. We analyze 758 English-language videos linked to major international
climate negotiation events and their associated comment sections. Using topic modeling to
examine video transcripts and sentiment analysis to study audience reactions, we identify
distinct patterns in how climate issues are framed and received. Videos that originate from
the Global North more frequently emphasize emission reduction policies and institutional
responsibility, while those from the Global South foreground development-related concerns.
Audience responses diverge more sharply: comment sections under Global North videos are
dominated by criticism and conspiracy-related discourse, whereas audiences are
comparatively more supportive and offer constructive arguments under Global South
videos. These findings highlight a gap between curated climate narratives and public
sentiment on YouTube and suggest that platform dynamics may reinforce or reshape
existing geopolitical divides in climate communication.

Abstract

Automatic Speech Recognition (ASR) for low resource Dravidian languages like Telugu and Kannada faces significant challenges in specialized medical domains due to limited annotated data and morphological complexity. This work proposes a novel confidence-aware training framework that integrates real and synthetic speech data through a hybrid confidence mechanism combining static perceptual and acoustic similarity metrics with dynamic model entropy. Unlike direct fine-tuning approaches, the proposed methodology employs both fixed-weight and learnable-weight confidence aggregation strategies to guide sample weighting during training, enabling effective utilization of heterogeneous data sources. The framework is evaluated on Telugu and Kannada medical datasets containing both real recordings and TTS-generated synthetic speech. A 5gram KenLM language model is applied for post-decoding correction. Results show that the hybrid confidence-aware approach with learnable weights substantially reduces recognition errors: Telugu Word Error Rate (WER) decreases from 24.3 % to 15.8 % (8.5 % absolute improvement), while Kannada WER drops from 31.7% to 25.4% (6.3% absolute improvement), both significantly outperforming standard fine-tuning baselines. These findings confirm that combining adaptive confidence-aware training with statistical language modeling delivers superior performance for domain-specific ASR in morphologically complex Dravidian languages.

Abstract

Conventional multirotor aerial vehicles actively suppress yaw rotation during hover, expending power to maintain a fixed heading despite the fact that yaw regulation is not required for force balance or altitude control. This paper challenges that paradigm by proposing a spinning quadrotor architecture that intentionally operates at a sustained yaw rate, converting power traditionally spent on yaw regulation into useful aerodynamic effects. A dynamic model of the spinning quadrotor is developed, analysis for low Re range is conducted to choose an airfoil for lifting surfaces. Preliminary hardware tests show a 22% reduction in thrust required. These findings suggest that intentional yaw rotation, rather than being suppressed, can be exploited as a design mechanism for efficient and robust multirotor flight.