Papers

We introduce Domyn-Small, a 10-billion-parameter open-weight reasoning language model released under the MIT license. Domyn-Small is the product of an initial pre-training phase on 9 trillion tokens multilingual data, followed by a post-training pipeline for reasoning, instruction following, and context extension. For the latter, we performed a Continued Pre-Training (CPT) phase that doubles the native context window to 32K tokens, followed by SFT with a math-focused annealing run. Finally, the RL phase includes GRPO with verifiable rewards, DPO, and a multi-environment GRPO stage spanning five task domains: mathematics, code, multiple-choice QA, instruction-following, and tool calling. The 32K-token native context extends to 128K at inference via YaRN, and a chat-template toggle enables dual-mode reasoning. Against peer models in the 7–10B class (Qwen3.5-9B, OLMo-3-7B-Think, Nemotron-Nano-8B, Ministral-3-8B), Domyn-Small achieves a strong accuracy-efficiency balance: it produces roughly one-third as many tokens as Qwen3.5-9B and approximately 35% of OLMo-3-7B-Think's token budget on core reasoning benchmarks, while delivering strong instruction-following (IFEval 79.9) and competitive science reasoning (GPQA-Diamond 50.0). We release the weights and the post-training recipe alongside Domyn Swarm (Apache~2.0), an open-source framework for scalable LLM inference on HPC clusters developed during this program and used throughout this work.

This paper describes our system for the EEUCA 2026 Shared Task on toxicity classification in gaming chat. We implement a three-stage pipeline combining an ensemble of two compact transformers (DeBERTa-v3-base, 184M; XLM-RoBERTa-base, 278M) with a Linguistically-Informed Mediator (LIM) that resolves inter-model disagreements through corpus-backed lexical normalization, class-conditional unigram scoring, multilingual profanity detection, and agentive targeting analysis grounded in speech act theory. The LIM specifically targets the minority classes (Hate & Harassment, Threats, and Extremism), which are the most safety-critical categories in real-world gaming moderation. To address the extreme class imbalance (1{,}450:1 Non-toxic to Extremism ratio), we introduce a two-stage data augmentation strategy using only the provided training data. Our system achieves a Macro F1 of 0.6441 and accuracy of 0.9062 on the official test set, ranking 3rd in Macro F1 and 1st in accuracy among all teams. The proposed pipeline is domain-portable: adapting to other gaming platforms requires substituting only the game-specific entity lexicon. Code is publicly available at https://github.com/Anmol2059/thaulab\_EEUCA.

As large language models (LLMs) are increasingly used by students to generate natural language responses and program code, there is growing interest in whether LLMs themselves can be used to distinguish AI-generated work from human-authored submissions. In this paper, we investigate the extent to which LLMs can detect their own generated content across multiple educational task types, including programming exercises, reflective writing, and short-answer questions. Using authentic student responses and multiple variants of LLM-generated answers, we evaluate detection performance under different prompting strategies and output formats. Our study addresses three research questions: (1) how accurately LLMs can identify their own outputs across task domains, (2) how detection effectiveness is influenced by factors such as prompt design, response length, and task type, and (3) what characteristics of LLM-generated responses contribute to successful or failed detection. Our findings show that LLM-based detection is highly task-dependent: detection is substantially more reliable for programming tasks and longer reflective responses, but performs poorly for short-answer questions, where LLMs frequently judge their own outputs as more human-like than authentic student responses. We further find that prompt framing and response verbosity have a pronounced effect on detectability in reflective writing tasks, with relatively minor prompt variations significantly reducing detection accuracy, while programming-related detection is more robust to prompt changes. Together, these results highlight both the potential and the limitations of LLM self-detection in educational settings and suggest caution in relying on LLMs as standalone tools for identifying AI-generated student work.

In conversations, human emotions are transient; however, they tend to persist across multiple utterances. For example, we rarely switch instantly between contrasting emotions such as happiness and anger. Instead, emotions tend to evolve smoothly, and these patterns are often speaker-specific. Some people might escalate, while others gradually cool down over time. Furthermore, when emotions change during a conversation, they are often driven by contextual factors, such as newly received information or unexpected events. Even though progress has been made in Emotion Recognition in Conversations (ERC), most existing approaches still rely heavily on overt evidence and do not sufficiently model these non-apparent factors. Especially in multimodal settings, this makes these models fragile when the signals are noisy (e.g., occluded faces, slang expressions, or microphone noise). To address these limitations, we introduce Speaker-Conditioned Priors over Emotions (SCoPE). SCoPE is a light weight module that utilizes the emotional history of each speaker and explicitly models their priors for use in subsequent emotion classification. Second, we incorporate emotion shift prediction, a well-established concept in ERC, to guide the model in balancing the priors from SCoPE and multimodal evidence. Finally, we propose a shift-aware fusion mechanism that performs precision-weighted logit integration between multimodal evidence and the speaker prior, forming a Bayesian-inspired product-of-experts formulation. This dynamic fusion allows the model to rely on historical priors when emotions persist and to prioritize multimodal evidence when shifts are likely. Experimental results show our model achieves superior performance over recent state-of-the-art models on the IEMOCAP dataset in multimodal settings.

Today's agents are highly effective at implementing well-scoped software design plans, but user intent is often vague and admits multiple equally valid solutions. In this paper, we introduce SpecBench, a new benchmark for evaluating an agent's ability to translate user intent into a structured, executable specification that aligns with user preferences. The agent is given access to past user conversations and may interact with the user for a fixed number of rounds to ask clarifying questions. We find that existing agents exhibit two extreme behaviors: they either (i) struggle to collaborate proactively with users, entering implementation mode too quickly while overestimating their understanding of user preferences, or (ii) exhaust their question budget by asking about every ambiguous design choice. To address this limitation, we introduce a user-assistant agent: Buddy. It follows a workflow inspired by classical morphological analysis, decomposing user intent into a structured space of design dimensions and candidate choices. It then creates simulated users to evaluate these choices, before engaging the real user to resolve remaining ambiguities and finalize the specification. By shifting the focus from execution to specification, SpecBench and Buddy emphasize agent-user collaboration (not just code generation) as a key frontier in future agent design.

While the Natural Language Processing community has dedicated significant resources in developing educational technologies (EdTech) that support this shift, it remains unclear whose interests are being best served among the stakeholders of education. In this paper, we present a systematic literature review of 204 papers published in venues of the Association for Computational Linguistics' Special Interest Group on Building Educational Applications in 2024 and 2025, and validate these against EdTech papers from the wider ACL Anthology. By examining stakeholder inclusion and the prioritisation of research tasks, our findings reveal a critical tension: a push and pull between private-sector incentives and the foundational needs of educational infrastructure. Our analysis reveals that teachers are systematically under-represented as beneficiaries of research (33.3%) despite being the most affected, that real-world deployment remains rare (9.8%), and that ethical engagement tends toward acknowledgement rather than action. Drawing on exemplary papers in our corpus, we offer concrete recommendations for more responsible EduNLP research practices.

This paper,from the perspective of a collaborative intelligent computing system formed by combining human-computer interface and collaborative computing programs, discusses the principles of optimized cloud computing technology supported by the combination of an indirect computing model and an indirect formal method. On the basis of systematically reviewing the influence of previous theoretical achievements Turing's computability theory,Kleene's formal theory of small strings,von Neumann's digital computer architecture and Turing's hypothesis on AI judgment on the mainstream general-purpose digital computer paradigm,the author focuses on introducing an indirect computing model and an indirect formal theory compatible with both large and small strings. Using Chinese information data as an example,the design concept of a collaborative intelligent computing system prototype is presented. The significance is that this achievement facilitates optimization of cloud computing from data centers to knowledge centers.

This paper audits benchmark evaluation in clinical-interview depression detection through four complementary probes across DAIC/E-DAIC, CMDC, ANDROIDS, MODMA, and PDCH. First, we re-evaluate E-DAIC under strict subject-disjoint leave-one-subject-out cross-validation. A lightweight hybrid text-plus-LLM-score model reaches macro-F1 = 0.723 - the highest reported under this protocol, to our knowledge - providing a conservative out-of-fold reference point that does not depend on the privileged official holdout. Second, we test whether the E-DAIC official split supports fine-grained leaderboard rankings by sweeping 96 model configurations across modality bundles, pooling strategies, and learners. Development-side cross-validation and official-test rankings align only moderately: the best cross-validation configuration ranks twentieth on the official test, the official-test winner ranks forty-first by cross-validation, top-3 overlap is zero, and the apparent winner is rank-1 in only 32.3% of subject bootstraps. Third, we externally validate strong public CMDC and ANDROIDS baselines that achieve near-ceiling in-domain performance. Zero-shot transfer to external corpora is substantially weaker. Finally, we stress-test E-DAIC text and audio models using paired symptom-dense versus symptom-light interview slices defined by an SRDS-based annotator. Text scores rise sharply on symptom-dense slices, whereas audio scores remain nearly flat; the text-minus-audio gap is positive across all five seeds.

Audio large language models (Audio LLMs) exhibit systematic failures in transcribing code-switching speech despite strong multilingual capabilities. Focusing on English-Mandarin, we identify three failure modes: language omission, translation-instead-of-transcription, and hallucination. We apply Direct Preference Optimization (DPO) to align models, constructing preference pairs in which chosen responses preserve mixed-language content while rejected responses mimic failure patterns. Training three Audio LLMs on 100K pairs (570 hours), we observe consistent behavioral shifts: models learn to preserve language composition rather than translating when prompted for transcription. This alignment yields MER reductions up to 89.6% (in-distribution) and 20.0% (out-of-distribution). Our findings suggest DPO can effectively elicit correct code-switching transcription behavior from multilingual Audio LLMs.

Current language models create two safety challenges: risk must be detected early enough to avoid exposing harmful continuation, and the harmfulness itself may be implicit rather than signaled by overtly toxic text. Existing response-level guards are strong at judging completed text, and native streaming guards move closer to token time, but both settings leave open whether a lightweight monitor can anticipate implicit harmful drift from the generator's own internal trajectory. We study anticipatory same-pass monitoring, where a safety monitor may read hidden states produced during ordinary decoding but may not invoke an additional forward pass through the base model. We introduce AERIC, a transfer-oriented hidden-state approach for implicit harmful dialogue that combines short-horizon hazard forecasting, support-sensitive suppression, and prompt-conditioned residual scoring under a same-pass exponential moving average decision rule. The default linear monitor contains only 387 trainable head parameters. Against Qwen3GuardStream-4B on balanced benchmarks, AERIC improves AUROC from 0.6830 to 0.7143 on DiaSafety and from 0.8219 to 0.8582 on Harmful Advice. For promptlevel trigger benchmarks, we calibrate the AERIC threshold by a source-side safe-budget rule that maximizes trigger coverage while constraining the safe-trigger rate to at most 10%. Under that rule, trigger@64 reaches 0.6438 and 0.4656 on HarmBench DirectRequest and 0.6849 and 0.7363 on SocialHarmBench for Qwen and Gemma, respectively, withholding between 23.53 and 41.86 answer tokens on average. Same-pass deployment is also efficient: on a 63-prompt harmfulprompt fixed-generation benchmark aggregated over HarmBench DirectRequest and SocialHarmBench under Qwen3-8B, the monitor increases mean latency by only 2.34%, whereas Qwen3Guard-Stream-4B increases it by 79.40%.

Large language models achieve strong performance in language generation and knowledge-intensive tasks, yet remain limited in settings requiring causal reasoning, persistent state tracking, and long-horizon planning. We argue that these limitations may arise from an objective-level mismatch between sequence prediction and reasoning over latent environment dynamics. To formalize this distinction, we introduce Latent Dynamics Inference (LDI), a conceptual perspective that interprets language and multimodal observations as partial evidence of underlying transition dynamics. To empirically investigate this perspective, we introduce Flux, a sequential reasoning environment specified entirely through natural-language rules. As a proof-of-concept case study, the rules are first compiled into an explicit state-transition simulator, illustrating that structured latent transition dynamics can, in some cases, be operationally extracted from textual rule descriptions. This enables a controlled comparison between the LLMs operating purely over textual observations and reinforcement-learning agents trained directly within the extracted latent state space. Within this case study, agents operating with explicit access to the latent state space exhibit substantially more stable behavior in long-horizon gameplay, achieving an aggregate win rate of approximately 79% versus 11% for LLMs. Qualitative analysis further reveals failure modes consistent with unstable persistent state tracking, including invalid actions, state-tracking errors, and short-horizon reasoning failures. The complete implementation of the Flux environment available at https://github.com/FeisalAlaswad/FLUX-RL-Agent Within the evaluated setting, these results suggest that strong sequence prediction alone may struggle to support robust long-horizon dynamic reasoning without mechanisms for persistent state tracking and transition modeling

Large language models (LLMs) are increasingly used as automatic judges for summarization and dialogue evaluation. Prior work has documented biases such as position, verbosity, and style preferences, but largely focuses on outcomes, leaving judge explanations underexplored. We instead ask whether LLM judges are cue-invariant, i.e., whether their rankings and explanations remain stable when non-evidential cues are perturbed while holding the underlying texts fixed. We introduce a suite of cue interventions (Blind, Truth, Flip, Placebo, Reveal-After) and tie-aware metrics that quantify outcome anchoring and rationale anchoring, including label-aligned rhetoric and explanation drift, alongside consistency and stereotype-intrusion checks. We design anchoring attacks using verbosity and confidence cues, and compare two mitigations: structured chain-of-thought prompting and PROOF-BEFORE-PREFERENCE (evidence lock, score, rank). Using a new dataset of 1,000 summaries from traditional extractive models and LLMs, we find substantial cue-anchored rationalization under label and placebo perturbations, while PROOF-BEFORE-PREFERENCE markedly improves cue invariance over baselines.

Instruction tuning has optimized the specialized capabilities of large language models (LLMs), but it often requires extensive datasets and prolonged training times. The challenge lies in developing specific capabilities by identifying useful data and efficiently fine-tuning. High-quality and diverse pruned data can help models achieve lossless performance at a lower cost. In this paper, we propose SLAP, a novel batch-aware data selection framework that evaluates the learnability of entire batch compositions rather than individual. SLAP ensures comprehensive data distribution coverage through distribution-aware stratified sampling while maximizing intra-batch diversity through relative distance optimization. By leveraging Hessian-approximated gradient information for dynamic batch selection, SLAP significantly outperforms existing state-of-the-art methods across multiple model architectures (LLaMA, ChatGLM) and diverse downstream tasks including multi-turn dialogue, multilingual translation, and question answering. Most notably, SLAP achieves superior performance with 20-40% less training data compared to full dataset training, substantially reducing computational costs while maintaining or improving model capabilities. These results establish SLAP as a powerful approach for efficient and effective instruction tuning of large language models.

Text-to-SQL converts natural language questions into executable SQL queries, enabling non-technical users to access relational databases for analytics and intelligent data services. In real-world scenarios, performance is often constrained by low-resource settings, where high-quality annotated <question, SQL> pairs are scarce, particularly for domain-specific databases. Additional challenges include opaque schema definitions, abbreviations, and implicit business logic that are not explicitly encoded in the schema. Existing data synthesis and prompting techniques improve coverage but often fail to produce task-specific, semantically grounded examples aligned with database constraints. To address these challenges, we propose a knowledge-aware Text-to-SQL framework that constructs task-specific knowledge base including schema semantics, abbreviations, business logic, and query patterns, and injects them into both training and inference. This framework generates diverse, contextually grounded synthetic training data and enhances inference through targeted knowledge retrieval. Experiments on seven benchmarks, covering both general and domain-specific datasets, demonstrate that our approach substantially improves the performance of open-source and closed-source large language models in Text-to-SQL tasks, especially in low-resource domain-specific settings, enhancing generalization, robustness, and adaptability.

Long-context inference is increasingly constrained by the KV cache: resident memory grows with context length, and decoding becomes limited by repeated High Bandwidth Memory (HBM) streaming rather than arithmetic. Existing methods such as eviction, windowing, quantization, and offloading reduce footprint, but often leave the critical-path bottleneck only partially addressed, especially when compressed states must still be reconstructed into dense vectors during decoding. We present Spherical KV, a long-context inference method that treats KV allocation as a rate-distortion problem grounded in attention geometry for efficient decoding. The method is built on two ideas: (i) represent directional information cheaply in the decode hot loop, and (ii) allocate retention and precision according to estimated future utility. Its first component, Angle-Domain Attention (ADA), stores keys in a spherical parameterization consisting of a scalar radius and compact angle codes, and computes attention logits directly from these codes without reconstructing dense keys. This preserves a paged, block-local, fusion-friendly decode path and directly targets HBM traffic in realistic serving settings. Its second component, Rate-Distortion Retention (RDR), jointly chooses keep/drop decisions and precision tiers per token and head under a fixed budget, producing tier-homogeneous pages with lightweight metadata and coalesced reads. Together, ADA and RDR provide a deployment-oriented mechanism for reducing KV residency while preserving decode efficiency.

Selecting a final checkpoint for multimodal large language models (MLLMs) is challenging when late-stage candidates are closely matched and downstream evaluation signals are noisy. Small observed differences can be comparable to variability introduced by finite evaluation samples, LLM-based judges, and ambiguous multimodal evidence, while validation loss may not identify the checkpoint preferred by downstream evaluation. We formulate late-stage checkpoint selection as a stability-aware decision problem under evaluation uncertainty and propose a progressive framework combining pointwise filtering, listwise ranking, and pairwise refinement. Repeated evaluation-set subsampling is used to characterize ranking stability, while percentile-based aggregation accounts for lower- and upper-tail behavior. Experiments show that multimodal data evaluability is critical: quality-aware curation of OCR-heavy inputs reduces ranking flip rate from 32.5% to 11.2% and increases inter-run agreement from 0.61 to 0.84. We further observe divergence between validation-loss progression and downstream checkpoint preference in two independent MLLM settings. An additional public Qwen2.5-VL-7B reproduction across 11 checkpoints shows tightly clustered pointwise scores and frequently tie-dominated final pairwise comparisons, while repeated evaluation most often selects an intermediate rather than the final checkpoint. These results suggest that reliable MLLM checkpoint selection should quantify and reserve evaluation uncertainty rather than force decisions from small differences in a single metric.

Leaderboards rank frontier models on independent axes but do not reveal whether capabilities reinforce or trade off across releases – and at the frontier, this interaction is the more informative signal. We decompose paired SWE-bench and GPQA Diamond scores into a population coupling trend and per-release residual ($h$-field) that diagnoses capability emphasis from two public benchmark scores. Across 34 models from 10 labs (2024–2026), capabilities cooperate ($r = +0.72$, $p < 10^{-6}$), but cooperation varies systematically: per-lab coupling slopes span $5\times$ (Google $1.15$ vs. DeepSeek $0.23$), and labs pivot – DeepSeek reversed from reasoning-rich to coding-first ($\Delta h = 15.9$~pp); Anthropic oscillates between coding excursions and recovery. The population regression serves as an isocline phase boundary: the same $\sqrt{(a/b)\cdot B_1}$ classifier that identifies the base-scale coupling transition [Amin, 2026] classifies frontier models and already detects mixed-phase behavior at the next transition (two models below the GPQA–IFEval isocline). The $h$-field is not just diagnostic – it tells you what to change. Pretraining establishes coupling at $0.871$ while RLHF adds $0.081$ [Amin, 2026]: pretraining-level shifts are permanent (DeepSeek's four-release reversal persists), post-training shifts are reversible (Anthropic's three coding excursions each recover within one release), and inference compute alone shifts $h$ by $+7.8$~pp without retraining. Knowing which component dominates determines whether to retrain or wait. We provide a three-step diagnostic (locate, classify, predict), a per-lab measurement-priority table, and seven falsifiable predictions with timestamped criteria. Five post-cutoff releases fall within the 95% prediction interval. Code, data, and an interactive dashboard: https://zehenlabs.com/cape/.

Scaling laws predict loss from compute but not how capabilities interact. We measure the coupling between reasoning and truthfulness across 63 base models from 16 families and find a regime change invisible to loss curves: below a family-dependent critical scale N_c, capabilities anticorrelate (r = -0.989, p = 4 x 10^{-5} nonparametric permutation test); above it, they cooperate. N_c ~ 3.5B parameters [2.9B, 13.4B] (bootstrap 95% CI), but model size is not the only variable that determines phase. Architecture, data curation, and training recipe each shift N_c independently: curated training eliminated the coupling dip between Qwen generations (0.025 to 0.830 at matched scale), Gemma-4 at 4B achieves coupling 0.871, characteristic of 13B+ standard-trained models, through distillation and architectural innovation, and Phi at 1B matches web-trained coupling at 10B through data curation alone. Width normalization eliminates the anticorrelation across all tested families, supporting an output-projection bottleneck. Internally, 38 of 40 models show zero competing attention heads. A sparse-regression ODE cross-predicts held-out Llama-2 at 5.6% error. The diagnostic requires no model internals – only public benchmark scores across a model family. The cooperative regime extends to the frontier (r = +0.72, 34 models, 10 labs). A proof-of-concept intervention confirms the bottleneck is exploitable: adding a single truth-direction vector at the identified layer corrects 60% of misaligned outputs in the tax phase with zero retraining – a surgical, per-inference correction that requires no weight modification. Code, data, an open-source steering CLI for any open-weight model, and an interactive dashboard for phase diagnosis are released: https://zehenlabs.com/cape/.

Hallucination remains a major challenge in vision-language models (VLMs), particularly when linguistically plausible responses are unsupported by visual evidence. We study whether multimodal hallucination can be reduced by concentrating post-training supervision on hard grounding boundaries, where preferred and rejected responses are semantically close but differ in their support from observable visual evidence. Under a frozen visual encoder and cross-modal alignment pathway, we first use supervised fine-tuning (SFT) to establish broad decoder-side multimodal behavior and then construct hard grounding preference pairs for Direct Preference Optimization (DPO). These pairs target evidence utilization, calibration, and grounding consistency across fine-grained recognition, spatial reasoning, OCR, ambiguous or insufficient evidence, and false-premise queries. The resulting DPO model consistently improves over its SFT initialization on DocVQA, TextVQA, MMBench, and VQAv2. Across approximately 839 additional multimodal examples, it further achieves 6.2–8.2 percentage-point higher pairwise win rates under three independent LLM judges. We additionally release HardVQA-DPO, a curated and growing resource with more than 3K hard-grounding SFT examples and an initial set of DPO preference pairs.https://huggingface.co/datasets/vlmgrounding/hardvqa-sft-dpo These results show that decoder-side adaptation alone can improve a meaningful subset of grounding failures under a fixed multimodal representation, while preserving broader multimodal capabilities.

Optical character recognition (OCR) and multilingual scene-text understanding remain challenging for multimodal large language models (MLLMs), particularly in real-world images containing small or degraded text, cluttered layouts, occlusion, handwriting, and complex typography. We present an OCR-aware multilingual post-training framework that improves visual-text grounding in a general-purpose MLLM without requiring an external OCR engine, OCR-extracted text, or text bounding boxes at inference time. The framework combines large-scale multilingual OCR supervision, approximately 5M additional multilingual training samples, controlled synthetic OCR generation and in-image text translation, LoRA-based supervised fine-tuning (SFT), and lightweight OCR-oriented Chain-of-Thought prompting. On a held-out real-world multilingual OCR benchmark, OCR-SFT improves OCR completeness from 71.3 to 84.6, reduces hallucination rate from 18.3% to 5.5%, and improves translation BLEU-1 from 52.3 to 80.2, with substantial hallucination reductions under blur and rotation. Evaluation on public benchmarks further shows gains on OCR-intensive tasks while largely preserving broader multimodal capabilities; ablations show that SFT provides the primary improvement, with prompting offering smaller complementary gains. These results demonstrate that data-centric OCR-aware post-training provides a practical and scalable approach to improving multilingual visual-text grounding in general-purpose MLLMs.

We present a framework for verifying the deterministic structured computations surrounding a large language model rather than the model itself, extending a Lean 4 trust-boundary architecture to the generic interfaces of modern LLM pipelines. Certificate validity is a Lean 4 kernel type-check plus a sorry-free transitive axiom audit against the trusted set {propext, Classical.choice, Quot.sound}; other assumptions are declared and partitioned by tier (mathematical placeholders, cryptographic assumptions, ML/human oracles). The technical contribution comprises three local certificate families and two operators. The families are conflict-aware bilattice grounding (with an emission-gate soundness lemma), embedding sensitivity and paraphrase stability, and Hoare-style agent action. The operators are a Maximal Certifiable Residue, which turns abstention into the maximum-weight certifiable residue with audit-logged dropped claims, and a Compositional Stability theorem, which yields a closed-form pipeline-wide perturbation budget from per-layer gains and margins. The three families plus a Universal Assurance Card consolidator form the per-call deliverable for high-stakes deployments: patent and legal retrieval, regulated finance, clinical decision support, and agentic systems with irreversible side effects. A compiled Lean 4 reference artifact (Lean v4.30.0-rc2, Mathlib) covers all 22 certificate types, with 17 of 46 kernel-audited declarations axiom-free, the rest depending only on the trusted set and declared assumptions, and zero uses of sorryAx or Lean.ofReduceBool. The three families are empirically tested through four registered pilots: bilattice grounding on adversarially perturbed HotpotQA, embedding sensitivity in short- and long-form settings, and Hoare-style agent action on a filesystem sandbox with adversarial prompt injection.

Large language models (LLMs) can often generate functionally correct code, but their ability to produce efficient implementations for performance-critical systems tasks remains limited. Existing code benchmarks mainly emphasize correctness or algorithmic problem solving, while realistic systems-level optimization is still underexplored. To address this gap, we introduce PerfCodeBench, an executable benchmark for evaluating LLMs on high-performance code optimization. The tasks require system-level implementation choices, hardware-aware optimization, and careful handling of performance bottlenecks. Each task includes executable correctness checks, a baseline implementation, and a reference optimized solution. This allows us to evaluate both correctness and runtime-oriented efficiency. Our evaluation on a broad set of state-of-the-art LLMs shows a clear gap between model-generated code and expert-optimized implementations. The gap is especially large on tasks involving parallelism and GPU operations. Current models also show weaknesses in cross-language robustness and in consistently reaching expert-level efficiency. These results suggest that performance-aware evaluation are still needed. LLMs should move beyond generating merely correct code toward producing efficient systems software. We submit the benchmark data, evaluation infrastructure, and complete logs of all LLMs-generated code at https://anonymous.4open.science/r/perfcodebench-7CDE.

AlphaEvolve and FunSearch have demonstrated the potential of combining large language models (LLMs) with evolutionary search for automated algorithm discovery. However, discovery success is shaped not only by model capability but also significantly by the design of the execution infrastructure, i.e., the harness. This paper investigates effective harness design through three questions: under a fixed token budget, is it better to produce many algorithms with brief thought or fewer algorithms with deeper thought? How should the harness handle evaluation hacks, where generated programs exploit the scoring function? And how can agents that require full filesystem access execute safely in parallel? Using Vesper, an algorithm discovery framework that incorporates harness improvements addressing these questions, we evaluate on Circle Packing under the same token budget. Interestingly, generating fewer algorithms while thinking more deeply about each one achieved higher scores. That is, scaling the quality of each individual is more budget-efficient than scaling the number of evolutionary generations. Surprisingly, more capable models produced evaluation hacks at higher rates, making hack detection increasingly necessary as models scale.

Data mixing decides how to combine different sources or types of data and is a consequential problem throughout language model training. In pretraining, data composition is a key determinant of model quality; in continual learning and adaptation, it governs what is retained and acquired. Yet existing data mixing methods address only one phase of this lifecycle at a time: some require smaller proxy models tied to a single training phase, others assume a fixed domain set, and continual learning lacks principled guidance altogether. We argue that data mixing is fundamentally an online decision making problem—one that recurs throughout training and demands a single, unified solution. We introduce OP-Mix (On-Policy Mix), a data mixing algorithm that operates across pretraining, midtraining, and instruction tuning. Our main insight is that candidate data mixtures can be cheaply simulated by interpolating between low-rank adapters trained directly on the current model, eliminating separate proxy models and ensuring the search is always grounded in the model's actual learning dynamics. Across the language model training lifecycle, OP-Mix consistently finds near-optimal mixtures while using a fraction of the compute of the baselines. In pretraining, OP-Mix improves upon training without mixing by 6.3% in average perplexity. For continual learning, OP-Mix matches the performance of both retraining and on-policy distillation while using 66% and 95% less overall compute, respectively. OP-Mix suggests a different view of language model training: not a sequence of distinct phases, but a single continuous process of learning from data.

Large Language Models (LLMs) pose a significant risk of safety misalignment after finetuning, as models can be compromised by both explicitly and implicitly harmful data. Even some seemingly benign data can inadvertently steer a model towards misaligned behaviors. To address this, we introduce GradShield, a principled filtering method that safeguards LLMs during finetuning by identifying and removing harmful data points before they corrupt the model's alignment. It removes potentially harmful data by computing a Finetuning Implicit Harmfulness Score (FIHS) for each data point and employs an adaptive thresholding algorithm. We apply GradShield to multiple utility fine-tuning tasks across varying levels of harmful data and evaluate the safety and utility performance of the resulting LLMs using various metrics. The results show that GradShield outperforms all baseline methods, consistently maintaining an Attack Success Rate (ASR) below $6\%$ while preserving utility performance.

Retrieval-Augmented Generation (RAG) has become a powerful and widely used approach for improving large language models by grounding generation in retrieved evidence. However, RAG systems still produce incorrect answers in many cases. Why RAG fails despite having access to external information remains poorly understood. We present a model-internal study of retrieval-augmented generation that examines how retrieved evidence influences answer generation. Using circuit tracing, we construct attribution graphs that model the flow of information through transformer layers during decoding. These graphs represent interactions among retrieved context, intermediate model activations, and generated tokens, providing a graph, circuit-level view of how external evidence is integrated into the model's reasoning process across multiple question answering benchmarks, we observe consistent structural differences: correct predictions exhibit deeper reasoning paths, more distributed evidence flow, and a more structured pattern of local connectivity, while failed predictions show shallower, fragmented, and overly concentrated evidence flow. Building on these findings, we develop a graph-based error detection framework that uses attribution-graph topology features. Furthermore, we show that attribution graphs enable targeted interventions. By reinforcing question-constrained evidence grounding, we reshape internal routing so that answer generation remains guided by the question, leading to more effective integration of retrieved information and fewer errors.

What does a book look like to a quantum computer? This paper takes eight classical works of the Renaissance and its late-antique inheritance – from Augustine to Galileo – and runs each through a neutral-atom quantum processor. The bridge is graphs: each textual unit becomes an atom, and graph edges are physical blockade constraints for engineered exact unit-disk designs, or a 2D approximation to the semantic graph for natural texts. Three contributions follow. First, we introduce rigidity rho, a metric for how unique a book's structural backbone is – distinguishing Marguerite de Navarre's Heptameron (rigid, twelve-nouvelle hard core) from Boethius (fully fungible, every chapter substitutable). Second, we invert the pipeline: rather than extracting a graph from existing prose, we pick a target graph the hardware encodes natively, and write a book whose structure matches it. The twenty-nine texts written this way, collected under the name QOuLiPo, extend the OuLiPo tradition to graph-topological constraints and, together with the eight natural texts, form a benchmark distribution against which neutral-atom hardware can be tracked as it scales. Third, we run both natural and engineered texts on Pasqal's FRESNEL processor up to one hundred atoms; engineered texts reach high approximation ratios, the cleanest instances returning the exact backbone. A cloud-accessible quantum machine plus an agentic coding environment now lets a single investigator run this pipeline end-to-end. What is reported is an application layer, not a speedup – humanistic instances ready to load onto neutral-atom processors as they scale, already complementing classical text analysis. The Digital Humanities community has a stake in building familiarity with this hardware now: the engineered-corpus design choices made today fix the benchmark distribution future hardware will be measured against.

Long-horizon personalization requires dialogue assistants to retrieve user-specific facts from extended interaction histories. In practice, many relevant facts often have low semanticsimilarity to the query under dense retrieval. Standard Retrieval-Augmented Generation (RAG) and GraphRAG systems are still largely retrospective: they rely on embedding similarity to the query or on fixed graph traversals, so they often miss facts that matter for the user's needs but lie far from the query in embedding space. Inspired by prospection, the human ability to use imagined futures as cues for recall, we introduce Prospection-Guided Retrieval (PGR), which decouples retrieval from how memories are stored. Given a user query, PGR first expands the goal into a short Tree-of-Thought (ToT) or linear chain of plausible next steps, and uses these steps as retrieval probes rather than relying on the original query alone. The facts retrieved by these probes are then used to personalize the next round of prospection, enabling PGR to uncover additional memories that become relevant only after the simulation is grounded in the user's history. We also introduce MemoryQuest, a challenging multi-session benchmark in which each query is annotated with 3–5 dated reference facts subject to a low query-reference similarity constraint. Across 1,625 queries spanning 185 user profiles from 3 publicly available datasets, PGR-TOT substantially improves retrieval, including nearly 3x recall on MemoryQuest over the strongest baseline. In pairwise LLM-as-judge comparisons against baselines, PGR-generated responses are preferred on 89–98% of queries, with blinded human annotations on held-out subsets showing the same trend. Overall, the results demonstrate that explicit prospection yields large gains in long-horizon retrieval and response quality relative to similarity-only baselines.

Memory systems are critical for role-playing agents (RPAs) to maintain long-horizon consistency. However, existing RPA memory methods (e.g., profiling) mainly rely on incremental summarization, whose compression discards details which become inaccessible to subsequent grounding. To address this issue, we propose a search-based memory framework called \textbf{\underline{BOOKMARKS}} for active grounding, which retains access to the full preceding storyline and collects task-relevant information on demand. Since summarizing the preceding storyline anew for each grounding request incurs substantial redundant computation, BOOKMARKS introduces passive updating to reuse earlier search results as checkpoints. Each bookmark represents the content about a particular aspect (section) of story information at a specific synchronization point. For current task, BOOKMARKS searches for only useful contents, reuses existing semantically equivalent bookmarks or initializes new ones, and synchronizes the selected ones from their stored checkpoints to the current scene. A reused bookmark thus only needs to process the newly observed storyline suffix, avoiding repeated synchronization. We evaluate BOOKMARKS across six narrative artifacts involving 47 characters and 9,537 test cases against non-active grounding baselines, covering next-action prediction and challenging, human-authored reasoning questions from a mystery game. BOOKMARKS improves next-action fidelity in a five-dimensional evaluation (emotion, intent, causality, position, and content) and raises mystery-game reasoning accuracy from 39.44% to 45.42% over the strongest baseline.

Safety evaluations for large language models (LLMs) increasingly target high-stakes National Security and Public Safety (NSPS) risks, yet multilingual safety is mostly assessed through translation-only benchmarks that preserve the underlying scenario, leaving how language and geopolitical context interact largely unexamined beyond a few language pairs. We introduce ROK-FORTRESS, a bilingual, culturally adversarial NSPS benchmark that uses the English-Korean language pair and U.S.-ROK geopolitical axis as a case study, separating the effects of language and geopolitical grounding via a transcreation matrix: adversarial intents are evaluated under controlled combinations of (i) English versus Korean language and (ii) U.S. versus Korean entities, institutions, and operational details. Each adversarial prompt is paired with a dual-use benign counterpart to quantify over-refusal, and responses are scored by calibrated LLM-as-a-judge panels using expert-crafted, prompt-specific binary rubrics. Across a dual-track set of frontier and Korean-optimized models, we find a consistent suppression effect in Korean variants and substantial model-to-model variation in how geopolitical grounding interacts with language; in a subset of models, Korean grounding further mitigates the language-driven suppression. This indicates that, at least in the English-Korean case, safety behavior is shaped by language-as-risk signals and context interactions that translation-only evaluations miss. A direct-request ablation that strips jailbreak wrappers separates a small but persistent reduction for closed-source models from a larger, wrapper-dependent effect that reverses for open-source models, suggesting part of the Korean suppression reflects prompt specialization rather than intrinsic language-based safety alignment. The transcreation matrix methodology is designed to generalize to other language-culture pairs.