| Total: 109
Advances in inference methods have enabled language models to improve their predictions without additional training. These methods often prioritize raw performance over cost-effective compute usage. However, computational efficiency is key for real-world applications with resource constraints. We provide a systematic analysis of the inference scaling strategies self-consistency, self-refinement, multi-agent debate, and mixture-of-agents, to study their computational performance tradeoffs. We evaluate methods on two reasoning benchmarks (MMLU-Pro, BBH) and include extensive parameter configurations (e.g., scaling the number of parallel predictions, agents, and debate rounds) across different model sizes. Across 34 configurations and over 100 evaluations, we compute the Pareto-optimal front to select methods that achieve the best accuracy with the lowest computational budget.Notably, inference scaling improves accuracy by up to +7.1% points over chain-of-thought at the highest evaluated budgets (20× the CoT compute budget) on MMLU-Pro. With an equal computing budget, debate and mixture-of-agents outperform self-consistency by 1.3% and 2.7% points, respectively. While self-consistency saturates earlier, multi-agent gains persist, particularly on more complicated tasks. We identify a simple multi-agent design guideline: mixture-of-agents is most efficient when the number of parallel generations exceeds the number of sequential aggregations.
The large and ever-increasing amount of data available on the Internet, coupled with the laborious task of manual claim and fact verification, has sparked interest in the development of automated claim verification systems. Several deep learning and transformer-based models have been proposed for this task over the years. With the introduction of Large Language Models (LLMs) and their superior performance in several NLP tasks, we have seen a surge of LLM-based approaches to claim verification along with the use of novel methods such as Retrieval Augmented Generation (RAG). In this survey, we present a comprehensive account of recent claim verification frameworks using LLMs. We describe the different components of the claim verification pipeline used in these frameworks in detail, including common approaches to retrieval, prompting, and fine-tuning. Finally, we describe publicly available English datasets for this task.
We investigate how multilingual representations emerge across depth in large language models.Using a unified probing framework, we analyze six multilingual LLMs across five languages (EN/ES/ZH/FR/DE), decomposing behavior into (i) early-layer dynamics, (ii) linear vs. MLP separability, and (iii) token–language alignment that tracks where vocabulary sharing peaks.Across models, we observe a consistent and substantial early jump: accuracy rises by +73.5 to +80.7 points from L0 to L1 on average, indicating that language-relevant signals become accessible immediately after the embedding layer.Moreover, representations are largely linearly separable: for 5/6 models, the mean gap between MLP and linear probes remains within ±0.5 points.Token–language alignment further reveals systematic structure, with peak vocabulary mass exceeding 48% in some models and substantial variation in the depth of peak sharing.These findings provide a compact, cross-model characterization of how multilingual information is organized across depth and introduce simple alignment metrics that complement accuracy-based evaluation.
Mechanistic interpretability aims to reverse-engineer large language models (LLMs) into human-understandable computational circuits. However, the complexity of pretrained models often obscures the minimal mechanisms required for specific reasoning tasks. In this work, we train small, attention-only transformers from scratch on a symbolic version of the Indirect Object Identification (IOI) task, a benchmark for studying coreference-like reasoning in transformers. Surprisingly, a single-layer model with only two attention heads achieves perfect IOI accuracy, despite lacking MLPs and normalization layers. Through residual stream decomposition, spectral analysis, and embedding interventions, we find that the two heads specialize into additive and contrastive subcircuits that jointly implement IOI resolution. Furthermore, we show that a two-layer, one-head model achieves performance comparable to that of a multi-head model by composing information across layers primarily through query-key interactions. These results demonstrate that task-specific training induces highly interpretable, minimal circuits, offering a controlled testbed for probing the computational foundations of transformer reasoning.
With the emergence of Large Language Models (LLMs), new methods in Information Retrieval are available in which relevance is estimated directly through language understanding and reasoning, instead of embedding similarity. We argue that similarity is a short-sighted interpretation of relevance, and that LLM-Based Relevance Judgment Systems (LLM-RJS) (with reasoning) have potential to outperform Neural Embedding Retrieval Systems (NERS) by overcoming this limitation. Using the TREC-DL 2019 passage retrieval dataset, we compare various LLM-RJS with NERS, but observe no noticeable improvement. Subsequently, we analyze the impact of reasoning by comparing LLM-RJS with and without reasoning. We find that human annotations also suffer from short-sightedness, and that false-positives in the reasoning LLM-RJS are primarily mistakes in annotations due to short-sightedness. We conclude that LLM-RJS do have the ability to address the short-sightedness limitation in NERS, but that this cannot be evaluated with standard annotated relevance datasets.
We propose a lightweight and single-pass uncertainty quantification method for detecting hallucinations in Large Language Models. The method uses attention matrices to estimate uncertainty without requiring repeated sampling or external models. Specifically, we measure the Kullback–Leibler divergence between each attention head’s distribution and a uniform reference distribution, and use these features in a logistic regression probe. Across multiple datasets, task types, and model families, attention divergence is strongly predictive of answer correctness and performs competitively with existing uncertainty estimation methods. We find that this signal is concentrated in middle layers and on factual tokens such as named entities and numbers, suggesting that attention dynamics provides an efficient and interpretable white-box signal of model uncertainty.
Automatic prompt optimization (APO) has emerged as a powerful paradigm for improving LLM performance without manual prompt engineering. Reflective APO methods such as GEPA iteratively refine prompts by diagnosing failure cases, but the optimization process remains black-box and label-free, leading to uninterpretable trajectories and systematic failure. We identify and empirically demonstrate four limitations: on GSM8K with a defective seed, GEPA degrades accuracy from 23.81% to 13.50%. We propose VISTA, a multi-agent APO framework that decouples hypothesis generation from prompt rewriting, enabling semantically labeled hypotheses, parallel minibatch verification, and interpretable optimization trace. A two-layer explore-exploit mechanism combining random restart and epsilon-greedy sampling further escapes local optima. VISTA recovers accuracy to 87.57% on the same defective seed and consistently outperforms baselines across all conditions on GSM8K and AIME2025.
Adapting Large Language Models to the medical domain remains an active area of research, with multiple strategies proposed to leverage annotated and unannotated data effectively. In this work, we propose a thesis outline to compare three common adaptation approaches—Instruction Tuning, Continual Pretraining, and Reasoning-oriented Training. We identify 5 dimensions to analyse: i) the interaction between the adaptation technique and the tasks; ii) the impact of the data size on the downstream performance; iii) the differences between datasets required by the three techniques; iv) the impact of the techniques given the model size; v) the impact of the techniques given the language.We construct an evaluation framework composed by 5 multilingual medical NLP tasks (named entity recognition, relation extraction, question answering, case report form filling, argument mining), spanning on 21 datasets in English, Italian, and Spanish, for a total of 61 combinations of language and sub-task.
Pretokenization is a crucial, sequential pass in Byte-level BPE tokenizers, yet little work has been done to optimize it for edge-side inference. Our proposed new implementation, Peek2, serves as a drop-in replacement for cl100k-like pretokenizers used in GPT-3, LLaMa-3, and Qwen-2.5. After breaking down and analyzing the logic of the original cl100k pretokenizer, we introduced a new pretokenization algorithm with linear time complexity and constant, trivial memory usage, suited for edge scenarios. Test results show that it increases microbenchmarking throughput by up to 2.48× and delivers a 1.14× improvement in overall throughput across the entire Byte-level BPE encoding process, depending on the dataset, while providing identical results as the baseline Regex-based tokenizer.
Annotator disagreement on tasks like natural language inference (NLI) reflects genuine linguistic ambiguity, yet most fine-tuning recipes treat every example as equally learnable. We ask whether this external signal of ambiguity predicts per-example learning dynamics under LoRA, the most widely used parameter-efficient fine-tuning method, and find that it does. Correlating annotation entropy (from ChaosNLI’s 100 labels per example) with per-example area under the loss curve (AULC) on SNLI and MNLI, the correlation is positive in all 25conditions tested (Spearman 𝜌= 0.06-0.43), with decoder-only models showing stronger correlations than encoders at matched LoRA rank. More strikingly, under LoRA contested examples exhibit un-learning: their gold-label loss increases during training, a pattern that is largely absent under full fine-tuning and IA3 in the encoder setting where matched comparisons are available, and that we also observe under LoRA on two decoder-only models. The effect survives partial-correlation controls and replicates across seeds and datasets. A preliminary noise-injection experiment is consistent with these findings.
Understanding idiomatic and figurative language in images remains a fundamental challenge for vision–language models, as it requires reasoning beyond literal image–text alignment. Although large pretrained models such as CLIP and BLIP-2 perform well on literal recognition, they consistently fail on multimodal figurative benchmarks, often favoring visually salient but semantically literal interpretations. We show that this failure arises from a systematic literal alignment bias rather than limited model capacity. Motivated by this observation, we reformulate multimodal figurative understanding as a contrastive semantic deviation problem, where figurative images must be distinguished from visually plausible literal alternatives. We introduce a parameter-efficient adaptation of CLIP using Low-Rank Adaptation (LoRA) with hard literal negative mining, achieving targeted reshaping of multimodal representations without full fine-tuning. Experiments on the IRFL benchmark across idioms, metaphors, and similes demonstrate substantial improvements over zero-shot CLIP, BLIP- 2, ensemble-based, and knowledge-augmented baselines. Finally, we introduce FIGMENT, a multilingual figurative grounding evaluation spanning five idiom-rich languages, and show that the adapted model generalizes across languages despite being trained exclusively on English supervision.
Chain-of-Thought (CoT) prompting is the dominant strategy for eliciting step-by-step reasoning in large language models, but its effect on code generation is poorly understood. We present a controlled 2×2 study of Qwen2.5-Coder-1.5B and DeepSeek-Coder-1.3B (each in base and instruction-tuned variants) on HumanEval, MBPP, and LiveCodeBench, plus scale-validation runs on Qwen2.5-Coder at 7B and 14B and a preliminary evaluation of CodeLlama-7B. We find that instruction tuning reverses CoT’s effect on small Qwen models: CoT improves the 1.5B base (+13.4pp, p<0.001) but significantly degrades the 1.5B instruct variant (-15.2pp, p<0.001). The reversal is sharply scale-bounded — it disappears at 7B (-0.6pp) and goes slightly positive at 14B (+2.4pp) — while CoT’s positive effect on base models grows monotonically with scale (+13.4 → +28.7 pp). DeepSeek-Coder-1.3B is insensitive regardless of regime. A direct token-count and truncation analysis shows the mechanism: at 1.5B, CoT inflates Qwen Instruct’s mean output length by 112 tokens and pushes 7.6× more generations into truncation, where Pass@1 is 0%; at 14B, the same prefix produces complete code well within budget. Layer-wise probing shows all four small models encode prompt type by Layer 1–4 (>90% accuracy) — universally, whether CoT helps or hurts — demonstrating that representation does not determine interpretation: the same internal signal drives divergent downstream behavior depending on training regime and capacity. Building on these mechanistic findings, we develop a probe-guided style router that, when trained per model on a labeled training split, selects among 12 prompt styles via a single 84 ms forward pass; it is statistically indistinguishable from the best fixed style in 7/8 settings and significantly outperforms CoT where CoT is most harmful (p=0.012, h=+0.40). Our results argue against applying CoT blindly to small instruct code models: its effect depends on architecture, training regime, and scale in ways that are mechanistically detectable from early-layer activations.
Diffusion language models generate text by iteratively denoising all tokens in parallel, but when and where their hidden states encode whether the output will be functionally correct remains unknown.We present the first probing study of DLM internals, training linear classifiers on hidden states to predict functional correctness.Across two models (LLaDA-8B, Dream-7B) and four tasks, we find that DLMs uniquely accumulate correctness signal across denoising steps (AUC gains of 0.08–0.11 on reasoning tasks), absent in single-pass AR decoding. However, step-0 signal reflects prompt difficulty rather than diffusion-specific computation. Signal emergence is task-dependent: structural tasks show flat profiles while reasoning tasks show gradual buildup. The two models exhibit distinct layer dynamics, with LLaDA concentrating signal in upper layers while Dream redistributes toward lower layers. We further show that probe confidence can identify likely failures, enabling selective generation that avoids 36–98% of wasted compute.
Retrieval-augmented generation (RAG) based on dense embeddings has become a dominant paradigm for text retrieval. However, many real-world applications require attribute-specific querying, where explicit values or properties must be extracted from text (e.g., symptoms in clinical notes or dosage values in medical reports). Dense retrieval handles paraphrastic variation well but often entangles multiple attributes within a single embedding, making value extraction difficult. Knowledge graphs (KGs), in contrast, support explicit attribute access but are brittle under linguistic and structural variation, leading to low recall.This thesis proposal aims to investigate the representational trade-off underlying these approaches. We study knowledge graph representations from an information-theoretic and optimal coding perspective, focusing on the tension between fine-grained factorization and compact canonicalization of concepts. Building on this perspective, we propose a query-driven framework for constructing and retrieving knowledge graphs from text, aiming to combine the robustness of dense retrieval with the explicit queryability of symbolic representations.
We introduce TokLens, an open-source toolkit for evaluating tokenizer quality across languages using six intrinsic metrics: fertility, characters per token, compression ratio, normalized sequence length, single-token retention rate, and cross-lingual parity. We evaluate 24 tokenizers from major LLM families across 15 typologically diverse languages and correlate these metrics with downstream performance. Our analysis reveals stark disparities: GPT-2 produces 56x more tokens per word in Japanese than in English, while newer tokenizers like Qwen2.5 and Gemma-2 reduce this gap to under 4x. No intrinsic metric predicts English benchmark performance after controlling for model size. However, on multilingual benchmarks (MMLU-ProX), linear mixed-effects models show that tokenizer metrics significantly predict per-language performance (STRR: 𝛽 = +5.7, z = 18.5, p < 0.001). A controlled experiment on the Qwen2.5 family further shows that languages with higher single-token retention rate exhibit steeper scaling slopes (𝜌 = 0.91, p < 0.001). These results indicate that tokenizer quality is significantly associated with multilingual LLM performance, though the evidence remains correlational and partially confounded with pretraining data composition.
One partner says “Fine” meaning resolution; the other hears surrender. The word is shared; the affective uptake is not. We formalize this as affective meaning divergence (AMD), the total-variation distance between interlocutors’ anchor-conditioned affect distributions. Building on speech-act theory, common-ground accumulation, and entropy-regularized game theory, we derive a logit best-response map whose dynamics undergo a saddle-node bifurcation: when 𝛽𝛼 > 4, a monotone increase in AMD-driven load produces an abrupt, hysteretic collapse of repair coordination. On Conversations Gone Awry (CGA-Wiki; N=652), derailing conversations exhibit critical-slowing-down (CSD) signatures across multiple levels: lexical divergence variance (p<0.001, d=0.36), AMD variance (p=0.001, d=0.26), and dialog-act repair variance (p=0.016, d=0.20), all significant after correction and stronger than toxicity and sentiment baselines. AMD provides a distinct temporal signature, with retrospectively measured variance peaking at the bifurcation point while toxicity variance peaks earlier, and is the only indicator grounded in the theoretical framework. Boundary-condition analysis on CGA-CMV (N=1,169) yields mixed but directionally consistent evidence.
Reasoning models frequently agree with incorrect user suggestions - a behavior known as sycophancy. However, it is unclear where in the reasoning trace this agreement originates and how strong the commitment is. We introduce sycophantic anchors - sentences identified via counterfactual analysis that commit models to user agreement. Across four reasoning models spanning three architecture families (Llama, Qwen, Falcon-hybrid) and 1.5B - 8B parameters, we analyze over 200,000 counterfactual rollouts and show that linear probes reliably detect sycophantic anchors (74 - 85% balanced accuracy), outperforming text-only baselines at high commitment levels -confirming they capture internal states beyond surface vocabulary. Regressors further predict commitment strength from activations (R2 up to 0.74). We observe a consistent asymmetry: sycophancy leaves a stronger mechanistic footprint than correct reasoning. We also find that sycophancy builds gradually during generation rather than being determined by the prompt. These findings enable sentence-level detection and quantification of model misalignment mid-inference.
Neural IR models achieve strong performance but remain difficult to interpret. We present NEAT-IR, a black-box analysis framework that explains ColBERT’s ranking behavior using 26 classical IR features (BM25, TF-IDF, IDF measures, positional signals). We analyze ColBERT through two complementary lenses: regression (predicting exact scores) and learning-to-rank (predicting relative order), evaluated on MS MARCO (48,250 query-passage pairs). Our key finding is a score-rank gap: classical features preserve ColBERT’s rankings nearly perfectly (NDCG@5 ≈ 0.99) yet explain only R2 ≈ 0.28 of score variance. Feature attribution reveals that regression and ranking models rely on distinct feature subsets: query-level IDF signals dominate score prediction, while document-matching features (BM25, cosine TF-IDF) drive ranking preservation. These findings suggest that ColBERT’s ordinal behavior on MS MARCO is largely recoverable from classical signals, while neural contributions primarily affect score magnitude. NEAT-IR enables practitioners to diagnose when neural rankers deviate from classical patterns, supporting interpretable model auditing and informed hybrid pipeline design.
Large Language Models have demonstrated strong multilingual fluency, yet fluency alone does not guarantee socially appropriate language use. In high-context languages, communicative competence requires sensitivity to social hierarchy, relational roles, and interactional norms that are encoded directly in everyday language. Bangla exemplifies this challenge through its three-tiered pronominal system, kinship-based addressing, and culturally embedded social customs. We introduce BanglaSocialBench, the first benchmark designed to evaluate sociopragmatic competence in Bangla through context-dependent language use rather than factual recall. The benchmark spans three domains: Bangla Address Terms, Kinship Reasoning, and Social Customs, comprising 1,719 culturally grounded instances written and verified by native Bangla speakers. We evaluate twelve contemporary LLMs in a zero-shot setting and observe systematic patterns of cultural misalignment. Models frequently default to overly formal address forms, fail to recognize multiple socially acceptable address pronouns, and conflate kinship terminology across religious contexts. Our findings show that sociopragmatic failures are often structured and non-random; for example, inappropriate addressing choices concentrate heavily in downward-hierarchy (Elder→Younger) and informal contexts. This reveals persistent limitations in how current LLMs infer and apply culturally appropriate language use in realistic Bangladeshi social interactions.
Hate speech detection is essential for maintaining healthy online communities. Large language models (LLMs) perform well on text classification, yet their decision strategies need to be better understood. While post-hoc rationales can justify individual decisions, they substantially increase inference cost and limit scalability in high-throughput settings. As another approach, we propose an extended rational inattention model that parameterizes linguistic noise and information processing cost, providing an interpretable behavioral framework for black-box LLM classifiers. Treating LLMs as rational decision-makers under information constraints allows us to estimate - from the observed classification behavior - the parameters that represent information processing cost and noise sensitivity. As a case study and using a hate-speech dataset spanning multiple noise environments, we evaluate four commercial LLMs and show that the introduced extended rational inattention model predictions closely match the observed performance across different noise levels. We further test the performance under various noise mechanisms and find that the inferred information cost parameters remain consistent while the noise parameters vary with the distortion mechanism. Overall, our introduced framework offers a cost-efficient and quantitative approach to derive interpretable indices of LLM moderation behavior and decisions, without additional rationale generation.
Adversarial perturbations in the context of large language models (LLMs) are subtle changes added to input data (i.e., images or text) that are designed to alter predictions or outputs of machine learning models. We introduce several novel visualizations using topological data analysis (TDA) (leveraging persistent homology) to characterize how adversarial perturbations act on text inputs, specifically, how sandbagging and code-injection attacksalter the geometric structure of attention heads in transformer models. By computing persistent homology metrics from attention maps across different model architectures (such as BERT, RoBERTa, ELECTRA, DistilGPT, etc.), we find that adversarial inputs alter higher-dimensional topological features (H1 loops and H2 voids) in ways that distinguish them from clean, non-adversarial inputs.
We investigate whether large language models (LLMs) can generate literal usage examples for Japanese multiword expressions (MWEs), whose literal readings are structurally low-frequency in available corpora.Prior work on MWEs has largely focused on detecting idiomatic usages in context, leaving literal usages underrepresented particularly for Japanese MWEs whose literal readings are rare and structurally diverse.Because literal readings are rarely attested in corpora, we design a lexicon-grounded setup that uses corpus non-literal usages as contrastive cues for controlled prompting. We evaluate the generated sentences using automatic literalness judgments and human literalness judgments, together with manual inspection.Our results show that providing contrastive non-literal information stabilizes literal generation and improves quality compared with prompts that include only literal information or no hints. In addition, we conduct an LLM-based understanding test that compares model predictions of literal and idiomatic plausibility with human judgments.The results indicate that the model aligns more closely with human judgments for idiomatic interpretations than for literal ones, highlighting the relative difficulty of modeling literal readings of MWEs.The study demonstrates that LLMs can complement existing resources by supplying frequency-independent literal examples and offers a controlled framework for examining contextual meaning understanding of Japanese MWEs.
We propose a novel approach to translating Japanese slides into English andto correcting their layout errors by utilizing multimodal LLMs with slide images and XML structures.Existing translation tools often suffer from layout errors after translationdue to text expansion during the translation process, causing text to overlap with figures or other items in slides and thereby reducing readability. To overcome this issue, our proposed framework introduces two steps consisting of (i) translating text fragments within the slide, and (ii) correcting layout errors by optimizing layout placement based on visual consistency. In step (ii), we empirically show that few-shot prompts are quite effective in layout error correction. Given that the optimal combination of steps (i) and (ii) varies depending on the slide layout, our method generates eight different layout candidates. Consequently, we introduce a third step that automatically selects the optimal output from these eight candidates.The experimental results showed that the proposed method outperformed baselines and achieved 4.1% layout error rate and over 80% model selection success rate.
Rap is a vocal style rooted in Hip-Hop culture, characterized by producing rhymes in synchrony with a rhythmic beat.This paper proposes a method for generating Japanese rap lyrics with a large language model (LLM) whose rhyming behavior is improved via reinforcement learning.We design a reward function that evaluates end rhymes between two generated bars and apply GRPO, a reinforcement-learning method, to encourage Japanese rhyming without using existing Japanese rap lyrics as training data.Experimental results show that, although output collapse is observed in some cases, GRPO increases the proportion of outputs that receive moderate or high human ratings on rhyme-related criteria.
To prepare for an uncertain future, organizations must continuously monitor emerging trends and early signals of change. The increasing availability of web-based textual data has boosted natural language processing (NLP) methods in strategic foresight, particularly in the scanning phase. While prior studies have extensively focused on the identification of signals in such data, considerably less attention has been paid to how these signals evolve over time and gain relevance as they become more visible. This study addresses this gap by examining whether tracking the temporal dynamics of signals can improve their assessment for strategic decision-making. Demonstrated on the use case of the European electric vehicle market, we find three dominant signal trajectories and show that burst dynamics tend to surface signal consolidation rather than the early detection of weak signals. The results indicate that foresight research should move beyond static, one-off analyses toward a dynamic temporal perspective capable of identifying signals at earlier stages of emergence.