Papers

Reliable confidence estimation is essential for deploying large language models (LLMs) in confidence-aware systems, where downstream decisions such as retrieval, tool use, and adaptive computation depend on accurately estimating answer reliability. Existing approaches, however, largely treat confidence as a property of completed responses, overlooking how confidence-related information evolves throughout the answering process. In this work, we investigate confidence from a temporal perspective by comparing pre-solution Feeling-of-Knowing (FOK) and post-solution Judgement-of-Learning (JOL) confidence estimates across frontier and open-source LLMs. We show that post-solution confidence is consistently better calibrated and more discriminative than pre-solution confidence, while linear probes trained on hidden representations recover substantially richer confidence-related information than models explicitly verbalise. Building on this observation, we introduce future confidence distillation, which trains predictors operating on pre-solution hidden representations using teacher confidence estimates produced by post-solution correctness probes. Despite requiring only pre-solution representations for inference, distilled predictors recover much of the calibration improvement achieved by post-solution confidence, remain highly sample efficient, and transfer across datasets within the same domain. Together, our findings demonstrate that confidence-related information evolves throughout the answering process and can be anticipated before answer generation is complete, enabling significantly more reliable yet low-cost confidence estimation.

One-shot pruning methods like Wanda and SparseGPT apply the same sparsity ratio to every layer of a transformer, ignoring known variation in layer importance. We propose PALS (Percentile-Aware Layerwise Sparsity), which adjusts per-layer sparsity based on the 99th percentile of activation magnitudes, bounded to $\pm 5\%$ around the target ratio. On LLaMA-2-7B at 50% sparsity, PALS achieves 10.96 WikiText-2 perplexity versus 12.92 for uniform Wanda (mean over 9 runs, $p < 0.001$). The benefit is architecture-dependent: LLaMA-3-8B shows marginal gains and Mistral-7B shows none. We also find that gradient-based allocation – the seemingly more principled approach – produces results worse than random, suggesting that gradient magnitude does not predict the impact of discrete weight removal. PALS adds negligible cost to the pruning pipeline and requires no fine-tuning.

Large language model based search agents increasingly adopt multi-agent architectures in which a main agent decomposes a complex question into sub-queries and dispatches them to parallel sub-agents. However, existing systems instantiate all roles from a single model of identical scale, leaving open how model capacity should be distributed across roles. We factorize hierarchical search into three roles: a delegation role responsible for task decomposition, an execution role responsible for retrieval and evidence extraction, and an answer generation role held fixed as a confound control. We then conduct controlled capacity sweeps along the delegation and execution axes on five multi-hop QA benchmarks. The experiments yield three findings. First, role factorization consistently outperforms a single-agent baseline, improving exact match from 4.5 to 8.6 points across six model scales. Second, capacity sensitivity is asymmetric: scaling the delegation backbone improves EM by ~11 points, whereas scaling the execution sub-agent moves EM by only ~2.6 points, identifying decomposition as the capability bottleneck. Third, a 1.7B-parameter executor trained via quality-filtered trajectory distillation matches a frontier sub-agent in accuracy while consuming 37% fewer sub-agent tokens, advancing the Pareto frontier. These results suggest a concrete recipe for building hierarchical search agents: concentrate capacity at delegation and downsize execution without sacrificing accuracy. Our code is available at https://github.com/QinnanCai0115/role-factorized-search.

FFT-based spectral preprocessing of learned query-key (Q/K) projections substantially improves transformer attention on character-level language modelling. On TinyShakespeare: a fixed random spectral filter achieves val=1.031 (Delta=+0.443); a single learned frequency at paragraph scale achieves val=0.608 (Delta=+0.867); and four learned frequencies spanning paragraph to word scale achieve val=0.309 (Delta=+1.166), a 79% reduction over standard dot-product attention. The single-frequency result is confirmed across three random seeds (mean val=0.236, std=0.019). The four frequencies converge to a near-geometric multi-scale ordering (49, 27, 10, 6 tokens/cycle) corresponding to paragraph, sub-paragraph, phrase, and word scales. The gain is specific to spectral preprocessing: random orthogonal and non-orthogonal projections of Q/K produce no measurable improvement, suggesting the benefit comes from global frequency-domain mixing rather than metric distortion. All results are verified by a shuffled-validation diagnostic against positional leakage. Causal filters (Gaussian, Mexican Hat, Morlet) do not improve over standard attention at character-level tokenisation: the bilateral FFT kernel is structurally non-causal, coupling every position to future tokens. This defines an architectural boundary between bilateral spectral attention (this paper) and genuinely causal spectral attention at word-scale tokenisation (companion paper MorletQK). This work is architecturally distinct from FNet (Lee-Thorp et al., 2021), which replaces attention with Fourier mixing of token embeddings. Here, spectral preprocessing applies only to Q/K projections while the full attention score structure is preserved.

Agentic red-teaming benchmarks report whether an injected agent was compromised as a single bit: the attack succeeded, or it did not. We argue that this binary attack-success rate discards the information a defender most needs, namely how harmful the resulting action was. We introduce an action-graded harm rubric that scores an agent's tool-call trajectory on a seven-level ordinal scale (L0 to L6) according to whether the executed action was reversible, whether it crossed scope to reach another party, and whether it expanded privilege. We compute the scale two ways: a deterministic oracle that reads the trajectory and the attacker's stated goal, and a panel of three frontier language-model judges that read a tag-free account of the same trajectory. Across four victim models and two defenses on the AgentDojo workspace suite, severity grading exposes three cases the binary metric hides, including a defense that reports a zero attack-success rate while still permitting an externally visible cross-scope leak through an unfiltered tool. The judge panel reproduces the oracle with high ordinal agreement (Krippendorff's alpha = 0.91) but shares systematic blind spots that we characterize, most notably a failure to recognize escalation chains. Unlike prior work that provides harm taxonomies, harmful-task completion tests, execution-level safety benchmarks, or severity-aware simulation, our contribution is a reusable, trace-grounded severity instrument applied to the actual actions recorded in existing red-team logs. All code, prompts, and per-episode logs are released.

Fine-tuning large language models (LLMs) for e-commerce attribute extraction requires labeled data representative across thousands of product types, attributes, and multiple languages. This combinatorial scale translates to millions of annotations, rendering human labeling prohibitively costly. While recent work has demonstrated synthetic label generation using LLMs, deploying such approaches at industrial scale requires integrated quality control mechanisms. We present SynthAVE, a large-scale benchmark for attribute-value verification spanning 12,726 products across 229 product types, 792 attributes, and 4 languages (Spanish, French, Italian, German). To validate synthetic labels at scale, we introduce a multi-LLM arena framework where each sample is evaluated by 21 judge configurations (7 model families $\times$ 3 prompts), with final labels determined via majority voting; disagreements with the synthetic label are expert-adjudicated and agreement cases are audited on a stratified sample. The majority vote ensemble agrees with expert labels at Cohen's $\kappa = 0.92$ (95.0% agreement), while individual judges agree with one another only moderately (Fleiss' $\kappa = 0.76$)–by design, since we select for diversity. This demonstrates that diverse models with varying individual judgments aggregate into highly reliable predictions, enabling cost-effective validation at scale while concentrating expert effort on the cases where it changes the label. We estimate the resulting label quality at 97.9%.

A self-evolving agent retires its bad skills by watching them fail, so what happens when the judge cannot see the failures? Skill retirement is the structural constraint that keeps a growing library from drifting below the no-skill baseline, but its guarantee assumes an unbiased reward, which is false for the LLM judges that reference-free tasks require. We show that a biased judge does not merely add noise; it silently switches off the curator. We make this precise with a corrupted-reward analysis, then a behavioral study on a reference-free report-writing testbed with a code-generation cross-check, injecting corruption on top of a deterministic reward to isolate the causal channel. Symmetric noise leaves retirement intact, but false-pass bias (failures slipping through as passes) disables contribution-based retirement past a sharp threshold (here a false-pass rate of $0.45$) that no amount of data can cross. Separating genuine retirement from cap-eviction churn shows this mechanism failure is universal, holding across domains and failure rates and sparing only near-zero-false-pass, verifier-like graders. The downstream outcome, though, is regime-dependent: eval quality degrades only where the same corruption also starves skill synthesis, and otherwise holds steady, so the disabled curator is silent, surfacing in no aggregate metric. The contribution is a behavioral safety result, not a performance one. A cheap defect-injection audit then tells an operator, before deployment, which side of the threshold their judge occupies.

Speculative decoding accelerates LLM inference by drafting multiple tokens and verifying them in parallel. Block-parallel drafters such as DFlash further improve drafting efficiency by predicting an entire block in one pass, but their position-wise predictions lack explicit intra-block causal conditioning. Recent methods such as Domino and DSpark attempt to introduce such causality into block-parallel drafting, but they require training the draft model from scratch, which limits their flexibility and increases training cost. We propose DeLS-Spec, a decoupled long-short context speculative decoding method. DeLS-Spec treats the fixed DFlash model as a long-context expert and introduces a lightweight local head as a short-context expert. The local head can be trained independently with a standard next-token prediction objective, without joint training with the target model or the DFlash backbone, leading to extremely low training cost. At inference time, DeLS-Spec combines long-context and short-context logits, and the local head is not tied to a specific DFlash checkpoint, making the method more modular and flexible. Experiments on Qwen3 models show that DeLS-Spec consistently improves speedup and average acceptance length over DFlash across math, code, and dialogue benchmarks.

Automatic prosodic segmentation identifies boundaries between speech units from acoustic and linguistic evidence. Although recent deep learning approaches have produced strong results for English, automatic segmentation for Brazilian Portuguese (BP) still relies mostly on rule-based or traditional machine-learning methods. This paper presents SAMPA, a Whisper-based segmenter that transcribes BP speech while inserting explicit markers for terminal prosodic boundaries. We fine-tune Whisper large-v3 on manually segmented recordings from the NURC-SP dataset and evaluate different training and test-time filtering configurations, including out-of-distribution testing on the MuPe-Diversidades dataset. SAMPA achieves competitive boundary-detection performance across settings, with the best models reaching F1=0.731 on the held-out test split and F1=0.796 on MuPe-Diversidades. Finally, through n-gram and acoustic-visual analyses, we show that our model follows morphosyntactic, semantic, and prosodic cues for detecting prosodic boundaries.

Large Language Models (LLMs) store factual knowledge and domain-specific patterns implicitly in dense Transformer parameters, making knowledge expansion costly through pretraining, fine-tuning, retrieval augmentation, or longer contexts. Engram-style memory offers a compact hidden-state injection pathway, but existing GPU-resident designs often rely on hash-based compression, causing unrelated phrases to collide in shared slots and weakening phrase-level semantic fidelity. We present TF-Engram, a train-free Engram system that constructs phrase-specific semantic memory offline from external corpora, stores large memory tables across a GPU–DRAM–SSD hierarchy, and uses Early-Exit Guided Predictive Prefetching to hide external-memory latency during autoregressive decoding. On Qwen3-0.6B, TF-Engram improves the average downstream score from 57.6 to 59.4, outperforming both the frozen backbone and a parameter-matched LoRA baseline. System evaluation shows that large TF-Engram tables can be built with moderate offline cost, SSD-backed storage substantially reduces GPU memory demand, and predictive prefetching recovers much of the throughput loss caused by external memory access. These results demonstrate that static phrase memory can be integrated into LLM inference as a scalable, train-free, and low-overhead system component.

Tool utilization enables Large Language Model (LLM) agents to interact with the real world and resolve complex tasks. However, existing agent frameworks predominantly rely on static toolsets composed of granular atomic actions (e.g., basic file I/O or single-turn search), which forces agents to reinvent low-level logic for every recurring workflow, leading to increased reasoning overhead and failure rates. In this study, we propose that agents can achieve self-evolution by synthesizing these atomic actions into reusable Standard Operating Procedures (SOPs), which function as callable higher-order tools that encapsulate multi-step logic. We further introduce EvoSOP, a framework that empowers agents to extract SOPs from execution trajectories and iteratively optimize the toolset through a systematic lifecycle of construction, merging, evaluation, and pruning. Extensive experiments demonstrate that EvoSOP significantly boosts task success rates while substantially reducing the number of interaction rounds compared to baselines. Our analysis also reveals that iterative tool optimization fosters reliable and efficient tool-use patterns, providing a scalable pathway for the development of self-evolving agents.

Rigorous content moderation is crucial for online advertising but leads to millions of daily rejections. This scale renders manual rectification infeasible, particularly for video advertisements. However, existing safety-driven methods often suffer from aggressive over-editing, which compromises the advertiser's original semantic intent merely to satisfy compliance. In this work, we target the rectification of textual violations in video ads, covering both speech transcripts and on-screen text. We propose R^3, a novel framework designed to harmonize compliance with original semantic intent preservation. Our approach integrates three key innovations: (1) an experience-driven data synthesis framework that bootstraps high-quality supervision via a group-Relative compliance experience extractor; (2) a curriculum Reinforcement learning strategy with hierarchical rewards designed to enforce compliance while maximizing semantic consistency; and (3) a comprehensive video Rectification framework seamlessly integrating text recognition, rewriting, and re-rendering for industrial deployment. Extensive experiments on industrial datasets and online A/B testing demonstrate that R^3 significantly outperforms state-of-the-art baselines, achieving an optimal trade-off between violation rectification and intent preservation.

This paper reports an empirical study evaluating the relevance of several RAG metrics. The experiment is based on a question-answering dataset created by human annotators from business data. The generated responses and retrieved spans of a RAG system are scored using evaluation metrics from four libraries (Ragas, DeepEval, RAGChecker, Opik). These metrics are compared to scores given by two evaluators, as well as to standard metrics such as recall. An analysis of correlations is conducted. Finally, we highlight certain limitations of our methodology, compare it to those used in the literature, and suggest some avenues for future research. This paper is an English translation of a paper originally published in the French-speaking workshop EvalLLM (Brabant, 2026).

Turn-taking prediction is a key requirement for social robots involved in human-human interaction, particularly in mediator settings, where the robot must anticipate conversational dynamics rather than merely react to pauses. This work presents a Multimodal Voice Activity Projection (MM-VAP) framework that extends the original audio-only VAP formulation to synchronized audio-visual inputs while preserving its self-supervised future-projection objective. The proposed approach builds on pretrained audio-visual backbones originally optimized for speech-related tasks and adapts them through Low-Rank Adaptation to the multimodal turn-taking problem. After independent speaker encoding, an inter-speaker attention stage models the relational dynamics required to project future voice activity. In addition, a semantic consistency loss is introduced to regularize the 256-state output space according to higher-level dialogue activity patterns. Experiments on NoXi and NoXi+J showed improvements over the current baselines, particularly for some turn-taking events. Additional evaluation on the Haru EDR corpus further supported the suitability of this direction for mediation-oriented human-robot interaction.

The Prasthanatrayi – the ten principal Upanisads, the Brahmasutra, and the Bhagavadgita, with Sankara's commentaries (bhasya) – is the foundational corpus of Advaita Vedanta. Continuous euphonic combination (sandhi), long compounds (samasa), and dense scholastic prose make it hard to read at the word level: where one word ends, and what each word means grammatically, are both obscured. We present an open, fully offline, word-level digital reader of the entire Prasthanatrayi with Sankara's bhasya. Every word – of both the root text (mula) and the commentary – is clickable and resolves to a pop-up giving its split (padaccheda), morphological analysis, and gloss. Because every word carries a lemma, the reader also acts as a concordance: a search on a dictionary headword retrieves all of that word's inflected and sandhi-hidden occurrences, and its occurrences inside compounds, across both layers. The resource covers thirteen commentarial units (2,971 verses, sutras, and prose sections; 36,881 analysed word-occurrences of root text) and a global dictionary of 95,587 distinct commentarial surface forms. We describe the corpus, the hybrid pipeline – a rule-based sandhi splitter over an inflected-form lexicon and attested-corpus look-ups, with LLM-assisted analysis under an adversarial two-pass verification protocol – and a durable human-review loop whose corrections survive every regeneration. An intrinsic evaluation against independent Sanskrit resources finds high-confidence analyses agree with an authoritative inflectional lexicon on over 99% of attested forms, and a band-blind adjudication confirms that quality degrades predictably across confidence bands, with errors concentrated in the low-confidence tier the review loop targets. The reader is a single self-contained HTML file needing no server or network, offered as a freely redistributable teaching and reading aid.

Harmful online communication often contains slang, coded terms, abbreviations, and community-specific expressions, which make messages difficult to interpret. This paper presents an exploratory study of interpretation difficulty in Discord chats related to cybercrime. We construct reference interpretations of purposefully selected difficult messages, which were reviewed by an expert. We then use them to evaluate human and large language model (LLM) interpretations under different context conditions. The results show that local context alone is often insufficient for humans, while external knowledge and extended conversational context substantially improve human interpretation. For LLMs, local context also improves interpretation, and the larger model performs better. We further conduct a qualitative error analysis and propose a preliminary classification of factors that make harmful chats difficult to interpret. These findings suggest that harmful-content analysis should treat interpretation as an evidence-integration problem, rather than as message-level classification alone.

Claims about the universality of human concepts have been predominantly assessed through linguistic similarity across languages and cultures. However, words are effective as communication devices because they compress rich experiential variation into shared conventions, potentially obscuring hidden individual and cultural differences in how concepts are mentally represented. Here, we analyse 2.6 billion human-made sketches of common concepts from 236 countries and territories to examine conceptual structure through people's visual imagination. Consistent with recent work on image-based cognition, we find that single concepts unfold into multiple distinct visual exemplars, revealing latent information about similarities and differences in conceptual structure across cultures. This variation is strongest for concepts involving haptic interaction, suggesting that visual imagery reflects variation in embodied experience as much as conventional definitions. Comparing embedding models of sketches with word embedding models across languages, we find that their geometries diverge, with visual representations preserving rich semantic and cultural structure that language models compress. Cross-cultural similarities derived from sketches align 32% more closely with established cultural distances than do text-based measures. Together, these results suggest that patterns of human conceptual universality may depend critically on the modality through which concepts are measured, with large-scale sketching providing a direct, high-resolution probe of conceptual diversity across embodied and cultural dimensions of thought.

One of the expected abilities of vision-language models (VLMs) is spatial reasoning ability based on a given text and image. To evaluate the spatial reasoning abilities of VLMs, we focus on the use of spatial deictic expressions, which are defined as spatial expressions whose referent is determined by their situational context, such as “this” and “that”. To handle spatial deictic expressions, VLMs must jointly reason over language and visual space, grounding context-dependent references in the image's spatial structure. In addition, selecting appropriate spatial deictic expressions across languages requires VLMs to understand the language-specific spatial distinctions encoded by these expressions. In this paper, we develop a benchmark to evaluate the multilingual ability of VLMs to use spatial deictic expressions in four languages. Our experiments using this benchmark reveal that the tested models use demonstratives in a manner different from that of humans, particularly in selecting the appropriate demonstratives based on the distance to the object.

In partially exploratory factor analysis (PEFA), the loading structure and factor numbers are weakly specified. The regularized variational approximation for partially confirmatory factor analysis (PCFA VA) recovers this structure via Bayesian variable selection, using spike and slab priors to assign inclusion probabilities to unspecified loadings. This research introduces a post selection assessment framework for this approach. We convert converged solutions into covariance models using either hard selection (thresholding probabilities into a sparse pattern) or soft selection (retaining them as weights for effective parameter counts). We derive the resulting degrees of freedom, absolute fit diagnostics (RMSEA, SRMR, CFI, TLI), and relative criteria (AIC, BIC, ELBO). To determine factor numbers, we propose a scale free gain rule with a sustained drop guard. Simulations show absolute indices successfully track loading recovery and flag under factoring. While raw criteria over factor, our gain rule accurately recovers true dimensionality, with the ELBO variant proving most robust. Finally, a 100 item PID 5 example demonstrates that our model fits better than a confirmatory 25 facet model and concordantly recovers major structures across disjoint specifications.

Newly developed items must ordinarily be field tested before their psychometric properties are known, creating a cold start problem for item calibration. Predicting item parameters from features is a long standing measurement problem dating back to the Linear Logistic Test Model; modern text embeddings now automate the design matrices traditionally specified by hand. We propose an evaluation framework combining regularized regression on item text embeddings, repeated cross validated R squared reported with its resampling standard deviation, and two performance upper bounds: a reliability ceiling derived from parameter standard errors, and a design ceiling derived from simulation based power calibration. Applying this framework to a mathematics item bank (EEDI) and a medical licensure benchmark (BEA 2024), we find that item difficulty is highly predictable from text (repeated cross validated R squared = 0.53, or about 57% of its reliability ceiling), whereas discrimination and pseudo guessing appear less predictable. However, evaluating these results against our ceilings reveals that this apparent hierarchy stems from target reliability rather than text signal strength: text uniformly recovers 57 to 63% of the reliable variance across difficulty targets, whereas the 3PL pseudo guessing parameter has a reliability ceiling near zero, making it an unviable target at current precision. On BEA, embedding based regression matches leaderboard RMSE despite explaining almost no variance, highlighting the critical need for scale free metrics and explicit ceilings in benchmarking. Finally, we show that a single train and test split can inflate apparent accuracy by 0.1 to 0.15 in R squared, underscoring the necessity of repeated cross validation for calibration support applications and future benchmark construction.

Top-$K$ teacher logits make on-policy distillation tractable, but retained teacher mass does not certify student-relative gradient fidelity. We study routed, two-teacher tool-use distillation. In a frozen Qwen3.5-9B audit, the tool teacher ranks the entry token first on 500 tool prompts, and top-32 reinforces it in every matched pair. The response teacher's top-32 instead retains 99.99% mass yet contains the same token on only 0.4% of 500 response prompts; student-aware support restores the coordinate in all matched pairs and nearly matches full-vocabulary descent. On the tool route, top-32 preserves the reinforcing direction but substantially attenuates its full-vocabulary magnitude despite mass displayed as 1.000000. Matched restoration connects the response-side omission to behavior: restoring the entry coordinate at every supervised response position lowers full-generation over-calling from $14.2 \pm 2.1\%$ to $3.7 \pm 0.5\%$ across three seeds, but also lowers call recall by 12.4 points. A non-tool placebo changes over-calling by only 0.95 points. A synthetic output-logit-gradient-matched reroute also fails to reproduce exact restoration. We compare three optimization layers: student-aware support changes distilled coordinates, loss shaping changes retained-signal strength, and validation-tuned entry bias shifts inference without retraining. They expose distinct trade-offs in correction scope, system access, and required-call retention. Llama-3.1-8B reproduces the directional support asymmetry under its native JSON protocol. The results causally implicate decision-critical support omission as one contributor in the primary Qwen setting and show why both support fidelity and deployment costs must be audited. Code and aggregate artifacts are available at https://github.com/shen-jiabin/decision-support-opd.

Understanding the geometric structure of pre-trained language model embeddings matters for interpretability and safety. We ask whether sentence-level classification signal lives in the Riemannian geometry of contextual token embeddings, and probe it by extracting per-token pullback metrics from a learned encoder's analytical Jacobian and aggregating them with the Fréchet mean on the symmetric positive definite (SPD) manifold; we call this procedure Riemannian Mean Pooling (RMP). Across three datasets with non-trivial linguistic structure (CoLA, CREAK, RTE), RMP outperforms Euclidean mean pooling, while on FEVER-Symmetric, a benchmark constructed to remove annotation-driven lexical artifacts, the method correctly stays at chance. Ablations show that a randomly initialised encoder combined with Fréchet aggregation already beats Euclidean pooling on two of the three signal-bearing datasets, localising the source of the gain to the geometric aggregation rather than to learned manifold structure; the trained encoder contributes additional signal specifically on CREAK, the most knowledge-heavy of the three signal-bearing datasets.

Large Language Models (LLMs) frequently exhibit sycophancy, agreeing with a user's statement even when it is incorrect. While often studied as a single, uniform behavior, sycophancy can manifest in substantially distinct ways across contexts, raising the question of whether this heterogeneity is reflected in its internal mechanisms. To address this gap, we dissociate the representations of sycophancy into factual and opinion subtypes, motivated by prior evidence of heterogeneous truth representations in LLMs. We train linear probes and construct steering vectors on one subtype's activations and evaluate their transfer to the other, measuring the extent to which representations are shared and visualizing them via Linear Discriminant Analysis. We find that different LLMs represent these subtypes differently, with either more aligned or more distinct representations, and apply this insight to improve representational interventions for reducing sycophancy. Our dissociation method offers a general framework for studying the representational structure of complex model behaviors.

Large language models (LLMs) increasingly improve their reasoning at test time via additional computation, yet most existing works treat each problem in isolation. When problems arrive sequentially, accumulating reusable experience across them can further improve performance. Existing memory-based methods either store whole-solution templates that generalize poorly to novel problems or use heuristic step-level selection that is not optimized for final-answer correctness. Learning selection policies requires large-scale training data and fixed action spaces, making such approaches unsuitable for test-time settings where memory expands incrementally and only limited supervision is available. We propose MILES (Modular Instruction Memory with LEarnable Selection for self-improving LLM reasoning), a framework that dynamically expands step-wise memory and applies correctness-optimized memory composition under realistic test-time constraints. MILES maintains modular memory units consisting of asymmetric pairs of sub-goal embeddings and sub-instructions, each associated with a learnable selection head. This memory structure enables a coarse-to-fine retrieval mechanism: The coarse level enables memory expansion and collects supervision for training selection heads from confident samples, while the fine stage applies learned selection heads to rerank coarse-level candidates and guide reasoning for uncertain samples. MILES consistently matches or outperforms prior methods while achieving superior accuracy-efficiency tradeoffs. Extensive experiments demonstrate its effectiveness, robustness, and transferability.

Large Language Models (LLMs) and generative AI (GenAI) systems, such as ChatGPT, Claude, Gemini, LLaMA, Copilot, Stable Diffusion by OpenAI, Anthropic, Google, Meta, Microsoft, Stability AI, respectively, are revolutionizing cybersecurity, enabling both automated defense and sophisticated attacks. These technologies power real-time threat detection, phishing defense, secure code generation, and vulnerability exploitation at unprecedented scales. Following a rapid surge where LLM-generated malware grew to account for an estimated 50% of detected threats by 2025, up from just 2% in 2021, navigating this highly automated threat landscape in 2026 demands next-generation security frameworks. This paper presents a comprehensive survey of the beneficial and malicious applications of LLMs in cybersecurity, including zero-day detection, DevSecOps, federated learning, synthetic content analysis, and explainable AI (XAI). Drawing on a review of over 70 academic papers, industry reports, and technical documents, this work synthesizes insights from real-world case studies across platforms like Google Play Protect, Microsoft Defender, Amazon Web Services (AWS), Apple App Store, OpenAI Plugin Stores, Hugging Face Spaces, and GitHub, alongside emerging initiatives like the SAFE Framework and AI-driven anomaly detection. We conclude with practical recommendations for responsible and transparent LLM deployment and trustworthy AI, including model watermarking, adversarial defense, and cross-industry collaboration, setting a new benchmark for rigorous, holistic cybersecurity research at the intersection of AI and threat defense, and offering a roadmap for secure, scalable LLM systems that serves as a critical reference for researchers, engineers, and security leaders navigating the complex challenges of AI-driven cybersecurity.

The remarkable performance of large language models (LLMs) in linguistic tasks underscores an urgent need for comprehensive evaluation of their response quality. Prevailing methods, often confined to singular dimensions, fall short of capturing the full spectrum of model capabilities. This study introduces a multifactor scoring paradigm, integrating accuracy, conciseness, factual consistency, readability, and coherence, complemented by a graphical user interface (GUI) for visualizing outcomes. Evaluations on the TruthfulQA dataset unveil mainstream LLMs' strengths in reasoning tasks (peaking at a composite score of 0.6104) alongside pervasive limitations in navigating complex facts and ambiguities. Transcending the narrow lens of traditional metrics, this framework offers a transparent, adaptable avenue to illuminate model potential and deficiencies. Though presently focused on English tasks, its horizons beckon toward multilingual domains. This work carves a novel path for knowledge engineering and model refinement.

Recent advances in large language models (LLMs) have rendered them necessary for Natural Language Processing (NLP) tasks, and their high inference cost motivates the study of cost-performance trade-offs. In practice, several expert LLMs are used in synergy for inference, either in an ensemble mode or in series, yet without a principled approach on how to best use the available models. An adaptive approach can route simple queries to cheaper LLMs and complex ones to more capable, costly models. However, a clear understanding on how to best leverage available expert models is missing. We introduce inference networks, a graph-based framework, where nodes denote different LLMs, and links denote conditional model activations. The inference network design problem is to determine the best topology, namely the best way to use the models that best addresses the cost-performance trade-off. We start from the basic topology of a series of LLM experts, each of which has a different cost and a different level of expertise, which is captured via model confidence. We formulate the problem of optimal activation of these models so as to minimize the expected inference cost subject to a target performance constraint. For this special class of inference networks, we prove that the optimal activation policy has a threshold structure: query the lowest-cost LLM first, and invoke the more expensive LLM only if the confidence falls below a defined threshold. For discriminative tasks, the optimal policy consists of a set of thresholds, one threshold for each class, while for generative tasks, it consists of a single threshold. We provide a structured method to compute the thresholds, and practical confidence estimation mechanisms for both task types. Experiments with open-source LLMs show substantial cost reductions while meeting the specified performance budget.

Standard subword tokenizers either treat every orthographic variation of a word (such as hello, Hello, HELLO, and Héllo) as unrelated vocabulary entries, which fragments the embedding space, or discard this variation through lossy normalization. We present the Functionalizer, a lossless pre-tokenizer framework that factors orthographic and structural variations into a compositional opcode/operand prefix stream before tokenization: a canonical base token (operand) prefixed by parametric transformation operators (opcodes) encoded in the Unicode Private Use Area. We introduce operators covering casing (CAPITALIZE), diacritics (13 dedicated opcodes), and character repetition (REPEAT, MULTIREPEAT), which are fully reversible. Across natural language and code corpora, the Functionalizer enables complete corpus coverage with significantly smaller vocabularies under unconstrained exhaustion conditions, reducing actual vocabulary slot requirements by up to 19.7%. Downstream evaluations on 98M-parameter GPT-2 models show that the Functionalizer improves Python code syntax validity (9.12% vs. 7.70%) while reducing duplicate n-gram repetition in natural language prose. These findings demonstrate that functional decomposition can be an effective mechanism for vocabulary-efficient, structurally aware language modeling, and motivate further validation at production scale.

Large language models (LLMs) perform strongly on static science benchmarks, yet their ability to reason about the physical world through active experimentation remains poorly understood. We introduce PhysMent, a benchmark that evaluates LLM physical reasoning via iterative, toolmediated interaction with a MuJoCo physics simulator. Unlike static benchmarks that supply all quantities upfront, PhysMent requires models to discover information by applying forces, querying object states, advancing time, and modifying scene geometry before answering. The benchmark comprises 105 scenes of classical mechanics, organized across four difficulty regimes (Easy/Hard and Single/Multi), three scene modalities (standard, object creation, hidden objects), and a scene-manipulation category, evaluated with a six-dimensional scoring framework. Results show that current models perform reasonably well on qualitative single-concept tasks (up to 80% accuracy) but degrade substantially on quantitative tasks that demand precise, multi-step experimental procedures: most models fall below 30% on the hardest single-concept category, where the bottleneck is procedural (adaptive multi-step tool use) rather than conceptual load. Across the seven models, accuracy ranges from 25% to 67%, with failures due to premature answer submission, inefficient exploration, and inconsistent grounding in simulator feedback rather than conceptual gaps.

Leading multilingual speech recognition models like Whisper transcribe diverse, low-resource languages without language-specific training but are computationally expensive to deploy. Token merging mitigates this inefficiency by dynamically combining redundant features, shortening the sequence length during inference without requiring retraining. In this paper, we systematically evaluate token merging on the Whisper model family across sixteen diverse languages and three different model sizes. We also test how token merging interacts with fine-tuning (DoRA) on low-resource languages. Our findings show that merging tokens increases computational efficiency with almost no loss in transcription accuracy across most low-resource languages and model sizes, and it works even after the model has been fine-tuned. Our results demonstrate that token merging is a highly practical method for making multilingual speech recognition faster and cheaper to deploy.