Papers

In many practical LLM deployments, a single guardrail is used for both prompt and response moderation. Prompt moderation operates on fully observed text, whereas streaming response moderation requires safety decisions to be made over partial generations. Existing text-based streaming guardrails commonly frame this output-side problem as boundary detection, training models to identify the earliest prefix at which a response has already become unsafe. In this work, we introduce StreamGuard, a unified model-agnostic streaming guardrail that instead formulates moderation as a forecasting problem: given a partial prefix, the model predicts the expected harmfulness of likely future continuations. We supervise this prediction using Monte Carlo rollouts, which enables early intervention without requiring exact token-level boundary annotations. Across standard safety benchmarks, StreamGuard performs strongly both for input moderation and for streaming output moderation. At the 8B scale, StreamGuard improves aggregated input-moderation F1 from 86.7 to 88.2 and aggregated streaming output-moderation F1 from 80.4 to 81.9 relative to Qwen3Guard-Stream-8B-strict. On the QWENGUARDTEST response_loc streaming benchmark, StreamGuard reaches 97.5 F1, 95.1 recall, and 92.6% on-time intervention, compared to 95.9 F1, 92.1 recall, and 89.9% for Qwen3Guard-Stream-8B-stric, while reducing the miss rate from 7.9% to 4.9%. We further show that forecasting-based supervision transfers effectively across tokenizers and model families: with transferred targets, Gemma3-StreamGuard-1B reaches 81.3 response-moderation F1, 98.2 streaming F1, and a 3.5% miss rate. These results show that strong end-to-end streaming moderation can be obtained without exact boundary labels, and that forecasting future risk is an effective supervision strategy for low-latency safety intervention.

Ultra low-bit quantization brings substantial efficiency for Transformer-based models, but the accuracy degradation and limited GPU support hinder its wide usage. In this paper, we analyze zero-point distortion in binarization and propose a Binary Weights & Ternary Activations (BWTA) quantization scheme, which projects tiny values to zero and preserves the accuracy of extremely low-bit models. For training, we propose Smooth Multi-Stage Quantization, combining a Levelwise Degradation Strategy and a Magnitude-Alignment Projection Factor to enable stable and fast convergence. For inference, we develop a BWTA MatMul CUDA kernel with instruction-level parallel bit-packing and comprehensive binary/ternary MatMul implementations for both linear and attention operators, allowing seamless integration across Transformer architectures. Experiments show that BWTA approaches full-precision performance for BERT, with an average 3.5% drop on GLUE and less than 2% drop on five tasks, and achieves comparable perplexity and accuracy for LLMs. In efficiency, it delivers 16 to 24 times kernel-level speedup over FP16 on NVIDIA GPUs, and 216 to 330 tokens/s end-to-end prefill speedup with lower memory footprint on LLMs. As an algorithm-hardware co-design, BWTA demonstrates practical, low-latency ultra-low-bit inference without sacrificing model quality.

Large language models (LLMs) struggle to accumulate evidence across multiple rounds of user interactions, failing to update their beliefs in a manner consistent with Bayesian inference. Existing solutions require fine-tuning on sensitive user interaction data, limiting their applicability in privacy-conscious settings. We propose AdaptFuse, a training-free framework that externalizes probabilistic computation entirely from the LLM, where a symbolic module maintains a Bayesian posterior over a discrete hypothesis set, while a frozen LLM contributes semantic reasoning via multi-sample Dirichlet aggregation. The two signals are combined through entropy-adaptive fusion, which automatically weights each source by its predictive confidence, shifting reliance from the LLM to the symbolic posterior as evidence accumulates. We evaluate across three domains: flight recommendation, hotel recommendation, and web shopping; on Gemma~2 9B, Llama~3 8B, and Qwen~2.5 7B. AdaptFuse consistently outperforms both prompting baselines and fine-tuned Bayesian Teaching models on all tasks, with accuracy improving monotonically over interaction rounds. These results demonstrate that principled inference-time algorithms can substitute for fine-tuning in personalized recommendation, without storing or training on sensitive user data.

Goal-oriented conversational systems require making sequential decisions under uncertainty about the user's intent, where the algorithm must balance information acquisition and target commitment over multiple turns. Existing approaches address this challenge from different perspectives: structured methods enable multi-step planning but rely on predefined schemas, while LLM-based approaches support flexible interactions but lack long-horizon decision making, resulting in poor coordination between information acquisition and target commitment. To address this limitation, we formulate goal-oriented conversation as an uncertainty-aware sequential decision problem, where uncertainty serves as a guiding signal for multi-turn decision making. We propose a Conversation Uncertainty-aware Planning framework (CUP) that integrates language models with structured planning: a language model proposes feasible actions, and a planner evaluates their long-term impact on uncertainty reduction. Experiments on multiple conversational benchmarks show that CUP consistently improves success rates while requiring fewer interaction turns. Further analysis demonstrates that uncertainty-aware planning contributes to more efficient information acquisition and earlier confident commitment.

LLM-based social simulations can generate believable community interactions, enabling “policy wind tunnels” where governance interventions are tested before deployment. But believability is not causality. Claims like “intervention $A$ reduces escalation” require causal semantics that current simulation work typically does not specify. We propose adopting the causal counterfactual framework, distinguishing necessary causation (would the outcome have occurred without the intervention?) from sufficient causation (does the intervention reliably produce the outcome?). This distinction maps onto different stakeholder needs: moderators diagnosing incidents require evidence about necessity, while platform designers choosing policies require evidence about sufficiency. We formalize this mapping, show how simulation design can support estimation under explicit assumptions, and argue that the resulting quantities should be interpreted as simulator-conditional causal estimates whose policy relevance depends on simulator fidelity. Establishing this framework now is essential: it helps define what adequate fidelity means and moves the field from simulations that look realistic toward simulations that can support policy changes.

Large language models (LLMs) often produce confident but incorrect answers, in part because standard evaluation incentives reward guessing over expressing uncertainty. We study epistemic abstention for factual questions with verifiable answers, where the goal is to improve selective answering, making LLMs abstain when they are likely to be wrong while preserving correct answers. Inspired by human behavioral decisions in question answering, we introduce I-CALM, a prompt-level framework for black-box LLMs. I-CALM combines elicited verbal confidence, announced answer/abstain payoffs, and normative guidance emphasizing truthfulness, humility, evidential support, and responsibility. To distinguish targeted abstention from indiscriminate refusal, we use a two-stage evaluation protocol, in which LLMs first choose whether to answer or abstain, and are then forced to provide a best guess for the abstained ones. Across models and factual QA datasets, I-CALM improves selective answering by reducing false-answer rate among surfaced responses and improving abstention quality, shifting error-prone cases into abstention while retaining answers the model would have answered correctly. Overall, I-CALM offers a lightweight way to improve inference-time selective answering without retraining or access to model's internal states. Code is available at https://github.com/FayLONG03/hallucinationControl.

A target-oriented proactive dialogue system is designed to steer conversations toward predefined targets while actively providing suggestions. The core paradigm of such a system is to plan a reasonable dialogue path and subsequently guide language models (e.g., pre-trained or large language models) to generate responses, where dialogue path planning serves as the central component-a novel yet under-explored problem. In this work, we propose a Forward-Focused Bidirectional Pseudo-Siamese Network (FF-BPSN) for dialogue path planning toward predefined dialogue targets. FF-BPSN employs two identical transformer-based decoders for forward and backward planning, together with a forward-focused module that integrates bidirectional information to construct the final forward path. This path benefits from bidirectional planning while prioritizing forward information. We then employ the planned path to guide language models in response generation. Extensive experiments on DuRecDial and DuRecDial 2.0 demonstrate that FF-BPSN achieves state-of-the-art performance in dialogue path planning and significantly enhances the effectiveness of target-oriented proactive dialogue systems.

Large language models (LLMs) have been increasingly used to analyze text. However, they are often plagued with contextual reasoning limitations when analyzing long documents. When long documents are processed sequentially, early or dominant concepts can overshadow less visible but meaningful interpretations, leading to cumulative analytical bias, omission error, and over-generalization. Additionally, independently generated outputs are often merged without systematic grounding, introducing redundancy, conceptual drift, and unsupported claims. This study proposes a structured framework combining parallel chunk-level processing with evidence-anchored consolidation. Texts are first divided into semantically coherent chunks and processed independently in parallel to remove influence from earlier processing. The independently generated interpretations are then consolidated using explicit evidence anchoring and prioritization that reduces dominance and over-generalization while improving traceability. Experiments with multiple model types and sizes indicate that parallel processing significantly reduces omission error by approximately 84%, increases evidence traceability by up to 130%, and reduces unsupported claims by up to 91%. Smaller models benefited most, suggesting that efficient parallel chunking and consolidation play a critical role in achieving reliable and scalable textual analysis.

Quantized Large Language Models (LLMs) are used more often in qualitative analysis because they run fast and need fewer computing resources. This study examines how different lower bits quantization levels (8-bit, 4-bit, 3-bit, and 2-bit) and quantization types affect the performance of LLaMA-3.1 (8B) on qualitative analysis. The study uses expert and non-expert responses from 82 interview transcripts. Low-bit models often produce higher levels of hallucinations and unstable results, especially when reading non-expert language with unclear terms. To improve performance, we propose a quantization-aware multi-pass prompt verification method. This method guides the model through controlled steps that reduce hallucinations. It removes unreliable content and passes the results to the next transcript after verification, improving accuracy. To validate performance, human coders analyzed transcripts using NVivo and BF16 LLaMA. BF16 LLaMA-3.1 produced high-precision output but had semantic drift and hallucination. These errors were corrected manually. The corrected BF16 output and NVivo human coding were combined to create a gold-standard ground truth (GSGT) for thematic extraction and frequency analysis. The results show that 8-bit models stay closest to the GSGT. The 4-bit models lose accuracy but become stable when the proposed method is applied. The 3-bit and 2-bit models drop in performance because of heavy compression, but they improve with the proposed prompt design and verification. The study also finds that models at the same bit level behave differently depending on quantization type. Overall, the method helps low-resource LLMs become more stable, accurate, and suitable for qualitative research at lower cost.

Decentraland, a decentralized virtual reality platform operating within the expanding Metaverse ecosystem, utilizes its native MANA token to facilitate virtual asset transactions and governance. This study investigates the integration of Discord community sentiment with multi-modal financial data to enhance cryptocurrency price prediction within virtual world economies. We address: (1) identifying sentiment patterns within Decentraland's Discord community, and (2) evaluating the impact of multi-modal features on token return forecasting. Using a BERT-based large language model for sentiment analysis, we develop two LSTM architectures: a baseline incorporating historical prices and a multi-modal variant integrating sentiment scores, trading volume, and market capitalization. Results indicate predominantly neutral community sentiment with a positive skew. The multi-modal model significantly outperforms the price-only baseline in prediction accuracy. These findings demonstrate the predictive value of community-derived signals for virtual economy forecasting and establish a foundation for future research at the intersection of immersive virtual environments, natural language processing, and cryptocurrency market analysis.

Diffusion language models (DLMs) are promising alternatives to autoregressive language models (ARMs), yet the intrinsic differences in their generated text remain underexplored. We first find empirically that off-the-shelf DLMs exhibit lower $n$-gram entropy, higher semantic coherence, and higher semantic diversity. To understand the cause, we conduct controlled experiments that decouple the effects of training objectives and decoding algorithms. Results suggest that the DLM training objective contributes to the increases in semantic coherence and semantic diversity, but has a minor influence on entropy. These differences are primarily driven by the bidirectional context; other components in the training objective, such as input masking, label masking, and the weighting function, have a much weaker influence. Further, our experiments demonstrate that the reduction in entropy stems from DLMs' decoding algorithms, particularly confidence-based remasking strategies. We provide a theoretical understanding for this entropy reduction phenomenon. Together, our work uncovers key mechanisms underlying the differences between DLMs and ARMs in text generation, and informs future design of training objectives and decoding algorithms in DLMs.

State-Space Models (SSMs) – structured SSMs (S4, S4D, DSS, S5), selective SSMs (Mamba, Mamba-2), and hybrid architectures (Jamba) – are deployed in safety-critical long-context applications: genomic analysis, clinical time-series forecasting, and cybersecurity log processing. Their linear-time scaling is compelling, yet the security properties of their compressed-state recurrent architectures remain unstudied. We present the first systematic treatment of SSM safety, security, and cognitive risks. Seven contributions: (1) Formal threat framework – SSM Attack Surface (five layers), State Integrity Violation (StIV), Cross-Context Amplification Ratio $\mathcal{X}_\mathcal{S}$, and a Spectral Sensitivity Proposition grounded in the $H_\infty$ norm. (2) Three novel attack classes: spectral adversarial attacks (transfer-function gain exploitation), delayed-trigger stateful backdoors (activate thousands of steps after injection), and state capacity saturation (entropy flooding forces silent forgetting). (3) 14 MITRE ATLAS technique extensions across the full tactic chain. (4) Six-profile attacker taxonomy with kill chains for genomics, clinical, and cybersecurity domains. (5) Four cognitive risk hypotheses grounded in state-compression mechanics. (6) Governance-aligned mitigations mapped to CREST, NIST AI 600-1, and EU AI Act. (7) Empirical evaluation: targeted genomic injection achieves $\mathrm{StIV}=0.519$ vs. $0.086$ random ($6.0\times$, $p<0.001$); PGD state injection achieves $156\times$ output perturbation over random; SSD-structured extraction confirmed at $O(N^2)$ vs. $O(N^3)$ query complexity ($N\times$ speedup). Validation on pretrained checkpoints is detailed in the Appendix.

Low-Rank Adaptation (LoRA) has become the dominant parameter-efficient fine-tuning method for large language models, yet standard practice applies LoRA adapters uniformly to all transformer layers regardless of their relevance to the downstream task. We introduce Aletheia, a gradient-guided layer selection method that identifies the most task-relevant layers via a lightweight gradient probe and applies LoRA adapters only to those layers with asymmetric rank allocation. Across 81 experiment rows covering 14 successful models from 8 architecture families (0.5B-72B parameters, including dense and Mixture-of-Experts architectures), with one additional documented failed Pythia/GPT-NeoX attempt in Campaign 2, Aletheia achieves a 15-28% training speedup (mean 23.1%, p < 0.001) with bounded extra forgetting and broadly matched downstream behavior on the evaluated MMLU, GSM8K, and HumanEval benchmark pack. Across the tested families and scales, Campaign 1 shows a 100% per-model speed win rate and Campaign 2 shows broadly preserved downstream behavior within a bounded-degradation framing. Together these results support a practical model-economics claim: intelligent layer selection can make LoRA fine-tuning materially more efficient without introducing major downstream damage on the evaluated set.

Existing evaluation methods for LLM-based AI systems, such as LLM-as-a-Judge, verdict systems, and NLI, do not always align well with human assessment because they cannot adapt their strictness to the application domain. This paper presents Temperature-Controlled Verdict Aggregation (TCVA), a method that combines a five-level verdict-scoring system with generalized power-mean aggregation and an intuitive temperature parameter T [0.1, 1.0] to control evaluation rigor. Low temperatures yield pessimistic scores suited for safety-critical domains; high temperatures produce lenient scores appropriate for conversational AI. Experimental evaluation on three benchmark datasets with human Likert-scale annotations (SummEval and USR) shows that TCVA achieves correlation with human judgments comparable to RAGAS on faithfulness (Spearman = 0.667 vs. 0.676) while consistently outperforming DeepEval. The method requires no additional LLM calls when adjusting the temperature parameter.

Recent advances in large language models (LLMs) have opened new opportunities for recommender systems by enabling rich semantic understanding and reasoning about user interests and item attributes. However, as privacy regulations tighten, incorporating user data into LLM-based recommendation (LLMRec) introduces significant privacy risks, making unlearning algorithms increasingly crucial for practical deployment. Despite growing interest in LLMRec unlearning, most existing approaches formulate unlearning as a weighted combination of forgetting and retaining objectives while updating model parameters in a uniform manner. Such formulations inevitably induce gradient conflicts between the two objectives, leading to unstable optimization and resulting in either ineffective unlearning or severe degradation of model utility. Moreover, the unlearning procedure remains largely black-box, undermining its transparency and trustworthiness. To tackle these challenges, we propose CURE, a circuit-aware unlearning framework that disentangles model components into functionally distinct subsets and selectively updates them. Here, a circuit refers to a computational subgraph that is causally responsible for task-specific behaviors. Specifically, we extract the core circuits underlying item recommendation and analyze how individual modules within these circuits contribute to the forget and retain objectives. Based on this analysis, these modules are categorized into forget-specific, retain-specific, and task-shared groups, each subject to function-specific update rules to mitigate gradient conflicts during unlearning. Experiments on real-world datasets show that our approach achieves more effective unlearning than existing baselines.

This paper presents PolySwarm, a novel multi-agent large language model (LLM) framework designed for real-time prediction market trading and latency arbitrage on decentralized platforms such as Polymarket. PolySwarm deploys a swarm of 50 diverse LLM personas that concurrently evaluate binary outcome markets, aggregating individual probability estimates through confidence-weighted Bayesian combination of swarm consensus with market-implied probabilities, and applying quarter-Kelly position sizing for risk-controlled execution. The system incorporates an information-theoretic market analysis engine using Kullback-Leibler (KL) divergence and Jensen-Shannon (JS) divergence to detect cross-market inefficiencies and negation pair mispricings. A latency arbitrage module exploits stale Polymarket prices by deriving CEX-implied probabilities from a log-normal pricing model and executing trades within the human reaction-time window. We provide a full architectural description, implementation details, and evaluation methodology using Brier scores, calibration analysis, and log-loss metrics benchmarked against human superforecaster performance. We further discuss open challenges including hallucination in agent pools, computational cost at scale, regulatory exposure, and feedback-loop risk, and outline five priority directions for future research. Experimental results demonstrate that swarm aggregation consistently outperforms single-model baselines in probability calibration on Polymarket prediction tasks.

Analogical reasoning is a core cognitive faculty essential for narrative understanding. While LLMs perform well when surface and structural cues align, they struggle in cases where an analogy is not apparent on the surface but requires latent information, suggesting limitations in abstraction and generalisation. In this paper we compare a model's probed representations with its prompted performance at detecting narrative analogies, revealing an asymmetry: for rhetorical analogies, probing significantly outperforms prompting in open-source models, while for narrative analogies, they achieve a similar (low) performance. This suggests that the relationship between internal representations and prompted behavior is task-dependent and may reflect limitations in how prompting accesses available information.

Black-box knowledge distillation for large language models presents a strict trade-off. Simple off-policy methods (e.g., sequence-level knowledge distillation) struggle to correct the student's inherent errors. Fully on-policy methods (e.g., Generative Adversarial Distillation) solve this via adversarial training but introduce well-known training instability and crippling computational overhead. To address this dilemma, we propose SODA (Semi On-policy Distillation with Alignment), a highly efficient alternative motivated by the inherent capability gap between frontier teachers and much smaller base models. Because a compact student model's natural, zero-shot responses are almost strictly inferior to the powerful teacher's targets, we can construct a highly effective contrastive signal simply by pairing the teacher's optimal response with a one-time static snapshot of the student's outputs. This demonstrates that exposing the small student to its own static inferior behaviors is sufficient for high-quality distribution alignment, eliminating the need for costly dynamic rollouts and fragile adversarial balancing. Extensive evaluations across four compact Qwen2.5 and Llama-3 models validate this semi on-policy paradigm. SODA matches or outperforms the state-of-the-art methods on 15 out of 16 benchmark results. More importantly, it achieves this superior distillation quality while training 10 times faster, consuming 27% less peak GPU memory, and completely eliminating adversarial instability.

The rapid deployment of open-source frameworks has significantly advanced the development of modern multi-agent systems. However, expanded action spaces, including uncontrolled privilege exposure and hidden inter-system interactions, pose severe security challenges. Specifically, Indirect Prompt Injections (IPI), which conceal malicious instructions within third-party content, can trigger unauthorized actions such as data exfiltration during normal operations. While current security evaluations predominantly rely on isolated single-turn benchmarks, the systemic vulnerabilities of these agents within complex dynamic environments remain critically underexplored. To bridge this gap, we systematically evaluate six defense strategies against four sophisticated IPI attack vectors across nine LLM backbones. Crucially, we conduct our evaluation entirely within dynamic multi-step tool-calling environments to capture the true attack surface of modern autonomous agents. Moving beyond binary success rates, our multidimensional analysis reveals a pronounced fragility. Advanced injections successfully bypass nearly all baseline defenses, and some surface-level mitigations even produce counterproductive side effects. Furthermore, while agents execute malicious instructions almost instantaneously, their internal states exhibit abnormally high decision entropy. Motivated by this latent hesitation, we investigate Representation Engineering (RepE) as a robust detection strategy. By extracting hidden states at the tool-input position, we revealed that the RepE-based circuit breaker successfully identifies and intercepts unauthorized actions before the agent commits to them, achieving high detection accuracy across diverse LLM backbones. This study exposes the limitations of current IPI defenses and provides a highly practical paradigm for building resilient multi-agent architectures.

Large language models are increasingly used for qualitative data analysis, but many workflows obscure how analytic conclusions are produced. We present QualAnalyzer, an open-source Chrome extension for Google Workspace that supports atomistic LLM analysis by processing each data segment independently and preserving the prompt, input, and output for every unit. Through two case studies – holistic essay scoring and deductive thematic coding of interview transcripts – we show that this approach creates a legible audit trail and helps researchers investigate systematic differences between LLM and human judgments. We argue that process auditability is essential for making LLM-assisted qualitative research more transparent and methodologically robust.

Can humans tell whether a news article was written by a person or a large language model (LLM)? We investigate this question using JudgeGPT, a study platform that independently measures source attribution (human vs. machine) and authenticity judgment (legitimate vs. fake) on continuous scales. From 2,318 judgments collected from 1,054 participants across content generated by six LLMs, we report five findings: (1) participants cannot reliably distinguish machine-generated from human-written text (p > .05, Welch's t-test); (2) this inability holds across all tested models, including open-weight models with as few as 7B parameters; (3) self-reported domain expertise predicts judgment accuracy (r = .35, p < .001) whereas political orientation does not (r = -.10, n.s.); (4) clustering reveals distinct response strategies ("Skeptics" vs. "Believers"); and (5) accuracy degrades after approximately 30 sequential evaluations due to cognitive fatigue. The answer, in short, is no: humans cannot reliably tell. These results indicate that user-side detection is not a viable defense and motivate system-level countermeasures such as cryptographic content provenance.

Large language models (LLMs) have been shown to encode truth of statements in their activation space along a linear truth direction. Previous studies have argued that these directions are universal in certain aspects, while more recent work has questioned this conclusion drawing on limited generalization across some settings. In this work, we identify a number of limits of truth-direction universality that have not been previously understood. We first show that truth directions are highly layer-dependent, and that a full understanding of universality requires probing at many layers in the model. We then show that truth directions depend heavily on task type, emerging in earlier layers for factual and later layers for reasoning tasks; they also vary in performance across levels of task complexity. Finally, we show that model instructions can affect truth directions; simple correctness evaluation instructions significantly affect the geometry and the generalization ability of the truth probes. Our findings indicate that universality claims for truth directions are more limited than previously known, with significant differences observable for various model layers, task difficulties, task types, and prompt templates.

Reverse engineering (RE) is central to software security, particularly for cryptographic programs that handle sensitive data and are highly prone to vulnerabilities. It supports critical tasks such as vulnerability discovery and malware analysis. Despite its importance, RE remains labor-intensive and requires substantial expertise, making large language models (LLMs) a potential solution for automating the process. However, their capabilities for RE remain systematically underexplored. To address this gap, we study the cryptographic binary RE capabilities of LLMs and introduce CREBench, a benchmark comprising 432 challenges built from 48 standard cryptographic algorithms, 3 insecure crypto key usage scenarios, and 3 difficulty levels. Each challenge follows a Capture-the-Flag (CTF) RE challenge, requiring the model to analyze the underlying cryptographic logic and recover the correct input. We design an evaluation framework comprising four sub-tasks, from algorithm identification to correct flag recovery. We evaluate eight frontier LLMs on CREBench. GPT-5.4, the best-performing model, achieves 64.03 out of 100 and recovers the flag in 59% of challenges. We also establish a strong human expert baseline of 92.19 points, showing that humans maintain an advantage in cryptographic RE tasks. Our code and dataset are available at https://github.com/wangyu-ovo/CREBench.

Large Language Models (LLMs) can compose poetry, but how far are they from human poets? In this paper, we introduce POEMetric, the first comprehensive framework for poetry evaluation, examining 1) basic instruction-following abilities in generating poems according to a certain form and theme, 2) advanced abilities of showing creativity, lexical diversity, and idiosyncrasy, evoking emotional resonance, and using imagery and literary devices, and 3) general appraisal of the overall poem quality and estimation of authorship. We curated a human poem dataset - 203 English poems of 7 fixed forms annotated with meter, rhyme patterns and themes - and experimented with 30 LLMs for poetry generation based on the same forms and themes of the human data, totaling 6,090 LLM poems. Based on POEMetric, we assessed the performance of both human poets and LLMs through rule-based evaluation and LLM-as-a-judge, whose results were validated by human experts. Results show that, though the top model achieved high form accuracy (4.26 out of 5.00, with Gemini-2.5-Pro as a judge; same below) and theme alignment (4.99), all models failed to reach the same level of advanced abilities as human poets, who achieved unparalleled creativity (4.02), idiosyncrasy (3.95), emotional resonance (4.06), and skillful use of imagery (4.49) and literary devices (4.67). Humans also defeated the best-performing LLM in overall poem quality (4.22 vs. 3.20). As such, poetry generation remains a formidable challenge for LLMs. Data and codes are released at https://github.com/Bingru-Li/POEMetric.

Large language models (LLMs) are increasingly being used for text classification across the social sciences, yet researchers overwhelmingly classify one text per variable per prompt. Coding 100,000 texts on four variables requires 400,000 API calls. Batching 25 items and stacking all variables into a single prompt reduces this to 4,000 calls, cutting token costs by over 80%. Whether this degrades coding quality is unknown. We tested eight production LLMs from four providers on 3,962 expert-coded tweets across four tasks, varying batch size from 1 to 1,000 items and stacking up to 25 coding dimensions per prompt. Six of eight models maintained accuracy within 2 pp of the single-item baseline through batch sizes of 100. Variable stacking with up to 10 dimensions produced results comparable to single-variable coding, with degradation driven by task complexity rather than prompt length. Within this safe operating range, the measurement error from batching and stacking is smaller than typical inter-coder disagreement in the ground-truth data.

Large language models (LLMs) excel at complex reasoning, yet their efficiency is limited by the surging cognitive overhead of long thought traces. In this paper, we propose LightThinker, a method that enables LLMs to dynamically compress intermediate thoughts into compact semantic representations. However, static compression often struggles with complex reasoning where the irreversible loss of intermediate details can lead to logical bottlenecks. To address this, we evolve the framework into LightThinker++, introducing Explicit Adaptive Memory Management. This paradigm shifts to behavioral-level management by incorporating explicit memory primitives, supported by a specialized trajectory synthesis pipeline to train purposeful memory scheduling. Extensive experiments demonstrate the framework's versatility across three dimensions. (1) LightThinker reduces peak token usage by 70% and inference time by 26% with minimal accuracy loss. (2) In standard reasoning, LightThinker++ slashes peak token usage by 69.9% while yielding a +2.42% accuracy gain under the same context budget for maximum performance. (3) Most notably, in long-horizon agentic tasks, it maintains a stable footprint beyond 80 rounds (a 60%-70% reduction), achieving an average performance gain of 14.8% across different complex scenarios. Overall, our work provides a scalable direction for sustaining deep LLM reasoning over extended horizons with minimal overhead.

Masked diffusion language models (dLMs) generate text through bidirectional denoising, yet this capability remains locked for infilling prompts. This limitation is an artifact of the current supervised finetuning (SFT) convention of applying response-only masking. To unlock this capability, we extend full-sequence masking during SFT, where both prompts and responses are masked jointly. Once unlocked, the model infills masked portions of a prompt template conditioned on few-shot examples. We show that such model-infilled prompts match or surpass manually designed templates, transfer effectively across models, and are complementary to existing prompt optimization methods. Our results suggest that training practices, not architectural limitations, are the primary bottleneck preventing masked diffusion language models from infilling effective prompts

Agentic search enables language models to solve knowledge-intensive tasks by adaptively acquiring external evidence over multiple steps. Reinforcement learning with verifiable rewards (RLVR) has emerged as a widely adopted training paradigm for search agents, yet outcome-only rewards are sparse and provide limited credit assignment for intermediate search actions. Existing process-reward methods therefore seek to densify supervision through proxy signals, external evaluators, or likelihood-based information gain. However, proxy rewards can deviate from the final outcome objective, while fixed evaluators can become stale as the search policy evolves, leading to unreliable process supervision. To address these challenges, we propose OASES, an Outcome-Aligned Search-Evaluation Supervision framework for agentic search. OASES derives outcome-aligned process rewards by evaluating how well each intermediate search state supports answering the original question. It further co-trains the search policy and the state evaluator on policy, allowing the evaluator to adapt to evolving search behavior and provide more reliable process rewards. Experiments on five multi-hop QA benchmarks show that OASES consistently outperforms strong RL baselines, with further analyses confirming the benefits of outcome-aligned process rewards and search-evaluation co-training.

Interpretability research has highlighted the importance of evaluating Pretrained Language Models (PLMs) and in particular contextual embeddings against explicit linguistic theories to determine what linguistic information they encode. This study focuses on the Italian NPN (noun-preposition-noun) constructional family, challenging some of the theoretical and methodological assumptions underlying previous experimental designs and extending this type of research to a lesser-investigated language. Contextual vector representations are extracted from BERT and used as input to layer-wise probing classifiers, systematically evaluating information encoded across the model's internal layers. The results shed light on the extent to which constructional form and meaning are reflected in contextual embeddings, contributing empirical evidence to the dialogue between constructionist theory and neural language modelling

Government agencies must register, classify and route every citizen appeal within statutory time limits, and much of this work is still done by hand. We describe AI Appeals Processor, a classification and routing component deployed in a CPU-only government environment, and report what its evaluation and deployment taught us. On 10,000 real Russian-language appeals from a cross-domain dataset, we compare Bag-of-Words and TF-IDF with SVM, fastText, Word2Vec+LSTM and multilingual BERT on a three-way appeal-type task. On a held-out test set of 1,500 appeals, BERT reaches 82% accuracy and Word2Vec+LSTM 78%, against 67% for individual operators measured on an expert-adjudicated gold standard. We deployed the LSTM: in a workflow where an operator verifies every prediction, its lower training cost made frequent retraining on operator-verified labels practical, while the four-point accuracy gap did not change the operator's task. End-to-end handling time fell by 53-56% across four appeal-length bands (unweighted mean 22.5 to 10.25 minutes); model inference takes under two seconds of this. Most residual errors trace to the label taxonomy rather than the model: the statutory definitions of complaints and applications overlap, and many appeals carry two intents. A post-deployment audit of production classifications, made after several retraining cycles by operators who saw the assigned category, judged more than 95% correct; we explain why this figure is not comparable with the test-set result.