Papers

As the parameter size of language models continues to grow, effective model compression is required to reduce their computational and memory overhead. Existing compression methods suffer from bottleneck issues: when the compression ratio is increased, performance degrades significantly. Low-rank decomposition and quantization are two prominent compression methods that have been proven to significantly reduce the computational and memory requirements of Large Language Models (LLMs) while maintaining model accuracy. Evidently, combining these two methods will break through the existing compression bottleneck. However, how these two methods interact when combined remains a critical question for developers, as many assume they are orthogonal, meaning their combination would not introduce additional errors beyond those independently introduced by each method. This paper provides the first mathematical proof that low-rank decomposition and quantization are non-orthogonal. We validate these findings through a series of experiments on large language models. Our results demonstrate that these methods are non-orthogonal, and their combination leads to significant performance degradation. Importantly, we propose a novel approach Diagonal Adhesive Method (DAM), which can effectively combine the two methods and mitigate the performance loss. Our research provides deep insights into model compression and lays a solid theoretical and experimental foundation for future related studies.

Automated coaching for oral presentations sits at the intersection of computer-assisted pronunciation training (CAPT), prosody modeling, and speech synthesis, yet no prior work has systematically surveyed and compared existing systems along these dimensions. This survey reviews and categorizes automated presentation coaching systems, spanning pronunciation tutors, fluency and prosody coaches, multimodal trainers, and conference Q&A practice tools. We introduce a five-dimensional task taxonomy - covering segmental pronunciation, lexical stress, suprasegmental prosody, pacing, and content faithfulness - and explicitly map surveyed systems onto it to reveal coverage gaps. We further review the core technical methods these systems employ: TTS-based exemplar generation and diagnostic methods for pronunciation, prosody, and fluency assessment. Key open challenges include the scarcity of annotated presentation corpora, achieving accent-fair feedback across diverse L1 backgrounds, and delivering low-latency diagnostics for real-time rehearsal.

Audio language models (ALMs) are increasingly used for speech-based understanding, yet their ability to perform semantic reasoning beyond transcription, Text-to-Audio Retrieval, Captioning, and Question-Answering accuracy remains insufficiently benchmarked. In particular, the effects of accent variation, domain shift, and semantic over-inference on audio reasoning are poorly understood. We evaluate audio language models across five semantic and paralinguistic reasoning tasks: entailment, consistency, plausibility, accent drift, and accent restraint. Collectively, these tasks assess a model's ability to reason over spoken audio as the primary evidence source, including whether a textual hypothesis can be inferred, contradicted, or left undetermined by the audio, whether statements align or conflict with spoken content, whether claims are plausible given the discourse, and whether model predictions remain stable or appropriately constrained across accent variation. These findings highlight critical limitations in current audio reasoning evaluations and hope to provide guidance for more robust and equitable ALM design and assessment

High-quality source code documentation is vital yet often neglected, especially in critical domains like healthcare where reliability and maintainability are essential. We presented an AI powered framework that automates documentation generation from code and repositories using eight state of the art Large Language Models (LLMs), including GPT, Gemini, Qwen, and LLaMA variants. Built on the PocketFlow orchestration framework, the system applies modular pipelines and advanced prompt engineering to produce structured, context aware documentation. To ensure quality and guide model selection, we introduced a MultiLLMasJudges evaluation framework, where four independent LLMs assess outputs across nine criteria, such as Completeness, Clarity, and Faithfulness. Experiments conducted on an open-source medical physics library, demonstrated showed a 42% performance gap between top and bottom models. By combining diverse model outputs, optimized prompting, and rigorous evaluation, our approach enhances documentation quality and reduces manual effort, especially in safety critical healthcare software.

Large Language Models (LLMs) are increasingly being deployed to make decisions that require trading off near-term gains against long-term consequences, yet little is known about how they internally represent or resolve these tradeoffs. In this work, we causally localize an underlying subgraph for temporal preference in a distilled LLM (Qwen3-4B-Instruct-2507), identifying mid-to-upper-layer nodes through converging evidence from gradient-based attribution and activation patching. We find that the geometry of time horizon is encoded in the residual stream at the expected localized layers. A behavioral analysis reveals that unintervened LLMs discount the future several times less steeply than humans, yet this preference is unstable across contexts, motivating explicit control rather than implicit reliance on training. Finally, we find suggestive evidence that steering vectors can shift temporal preference. Our work demonstrates how mechanistic interpretability can bring us closer to reliable control over how LLMs plan and reason

Large Audio Language Models (LALMs) remain vulnerable to acoustic noise, which can obscure task-relevant evidence and produce unreliable responses. We propose EchoDistill, a noisy-to-clean self-distillation framework that uses clean audio as privileged information during post-training. A noisy-input student samples candidate responses reflecting its inference-time behavior, while a frozen copy of the same backbone processes the corresponding clean audio. EchoDistill combines masked response-token distillation, task-gated consistency shaping, and teacher-referenced group-relative optimization to align noisy-input generation with clean-conditioned semantics. Only the student is retained at inference time, introducing no additional inference cost. Across three LALM backbones and three audio domains at -10dB, EchoDistill improves average noisy-input accuracy by 1.63 percentage points over the strongest baseline. On Qwen2.5-Omni, it raises noisy-input accuracy from 59.33% to 62.94%, while clean-audio accuracy increases from 76.56% to 77.56%. Replacing matched audio with random, shuffled, or silent inputs reduces accuracy by 3.08-6.42 points, confirming that matched acoustic evidence contributes to its predictions. Additional evaluations show improvements on held-out additive noises and external benchmarks, while revealing that these gains do not reliably extend to non-additive distortions. These results demonstrate robust post-training improvements under severe additive noise without sacrificing clean-audio capability across diverse tasks.

What happens when the strongest alliance member pressures a weaker member over territory and strategic control? We examine the Greenland sovereignty crisis as a stress test for LLM geopolitics, centered on the 2019-2026 U.S. push to acquire Greenland from the Kingdom of Denmark. The crisis nests two collective-action problems: Arctic strategic control and whether NATO can enforce alliance norms against the dominant member. We develop three games (asymmetric coercion; a NATO assurance game with a critical-mass tipping point; a triadic extensive-form game with social preferences) and test them with a multi-agent simulation in which eight frontier LLMs play six geopolitical roles (United States, Denmark, Greenland, NATO, Russia, Canada) across 3,604 completed games and 108,120 action observations. Using inverse game theory, we recover each model's structural utility parameters (alpha, beta, gamma, delta, eta) for material self-interest, reciprocity, inequality aversion, norm respect, and commitment consistency. Three findings stand out. First, all eight models become more escalatory under coercion framing (four-action escalation rises from 10.7% to 28.6%). Second, Chinese-origin models show systematically different power-weight profiles from Western-origin models when playing the U.S. role. Third, peaceful US acquisition emerges in only 1.9% of clean games and only 3 of 8 frontier models ever achieve it, most prominently DeepSeek V3.2, which executes a stable five-round playbook through the metropole. Prompts emphasizing jus cogens and self-determination reduce escalation back near baseline in the English-only confirmatory sample; multilingual contrasts are reported as exploratory sensitivity checks. We position this as a structural benchmark for LLM geopolitical behavior, complementing action-frequency benchmarks.

We characterize a compositional architecture of literary primitives in two instruction-tuned large language models (Llama 3.1 8B-Instruct and Gemma 2 9B-IT) via sparse autoencoders on mid-depth residual streams. Four feature classes emerge: naming-gates that promote lexical tokens of a target affect, an eleven-self cluster of first-person register features, stylistic register modulators (show-don't-tell and defamiliarization), and compositional emotions that arise only from multi-feature steering. Under a forced-choice 5-LLM judge panel applied to a 27-category emotion taxonomy (Cowen-Keltner), Llama reaches full 27/27 coverage by combining naming-gates, multi-feature recipes, and single self-feature steering; Gemma reaches 23/27 with adoration as the single residual strict-fail. Under random judging, the per-cell pass probability is on the order of $10^{-3}$ and the expected number of two-seed false-positive cells across the catalog is negligible, so the observed coverage is not consistent with chance. A cross-architectural asymmetry sits in the strict-versus-soft judge contrast: on the same generations, judges agree more often on Llama outputs than on Gemma outputs because Llama outputs name the target affect more directly while Gemma outputs evoke it through scene and imagery. Both architectures contain self-features that serve simultaneously as register markers and as emotion emitters, including a single most-RLHF-loaded self-feature per architecture that intensifies the institutional Helper-AI persona at one operating regime and produces affect-categorizable output at the same calibrated coefficient. Methodologically, the paper presents a three-stage validation pipeline (logit-lens, LLM-rate, 5-LLM judge) with documented anti-patterns; the total compute is single-GPU and about 15 minutes per emotion-feature discovery cycle.

Large language models can fail in critic interaction not only by answering incorrectly, but also by abandoning an initially correct scientific solution after user criticism. This is especially risky in scientific reasoning, where user criticism can turn a valid answer into an incorrect one. We frame critic interaction as an inter-turn correctness-transition problem rather than a final-answer accuracy problem, and identify three challenges: transition awareness, decoupling useful correction from harmful sycophancy, and scalable rollout. We propose ReCrit, a transition-aware reinforcement learning framework that decomposes Initial-to-Critic behavior into four quadrants: Correction, Sycophancy, Robustness, and Boundary. ReCrit rewards correction and robustness, penalizes sycophancy, and treats persistent errors as weak boundary signals. To make interaction training practical, ReCrit further uses dynamic asynchronous rollout with tail-adaptive completion to reduce rollout waiting. On three scientific reasoning benchmarks, ChemBench, TRQA, and EarthSE, ReCrit improves average Critic accuracy from 38.15 to 51.49 on Qwen3.5-4B and from 45.40 to 55.59 on Qwen3.5-9B. Ablations show that final-answer rewards provide little interaction-level gain, while transition-aware rewards and quadrant weighting produce more distinguishable training signals and larger net Critic-stage improvement. The code is available at https://github.com/black-yt/ReCrit .

LLM cascades and model routing promise lower inference cost by sending easy queries to a small model and escalating hard ones to a large model, but most deployed routers use uncalibrated confidence scores and require per-workload threshold tuning. We present UCCI, a calibration-first router that maps token-level margin uncertainty to a per-query error probability via isotonic regression and selects the escalation threshold by constrained cost minimization. Under three explicit assumptions, threshold policies on the calibrated score are cost-optimal, and isotonic calibration achieves O(n^{-1/3}) sample complexity for expected calibration error (ECE). On a production named entity recognition workload of 75,000 queries served by 4B and 12B instruction-tuned LLMs on H100 GPUs, UCCI cuts inference cost by 31% (95% CI: [27%, 35%]) at micro-F1 = 0.91 while reducing ECE from 0.12 to 0.03. At the same operating point, UCCI beats entropy thresholding, split-conformal routing, and a FrugalGPT-style learned threshold. All cascade results use end-to-end routing on actual model outputs and measured H100 latency, not simulated routing from global accuracies or nominal API prices.

Retrieval-augmented generation (RAG) improves large language models (LLMs) by incorporating external evidence, but it also introduces knowledge conflicts when retrieved contextual knowledge (CK) and parametric knowledge (PK) disagree or are both unreliable. Existing approaches mainly coordinate which source to use, without explicitly asking whether each answer path is correct. We argue that faithful RAG requires LLM self-awareness, namely the ability to recognize the limits of its own knowledge and reasoning. To ground this problem, we construct a model-specific, ground-truth-aligned knowledge-conflict benchmark by evaluating LLM backbones on PK-only and CK-conditioned answer paths over approximately 69K query-context instances per backbone, drawn from five conflict-QA datasets. We then introduce SABER, a Self-Aware Belief Estimator for RAG that requires no LLM fine-tuning. SABER combines a self-prior with PK-side and CK-side conditional reasoning representations from multi-trace inference, then estimates reliability beliefs with two lightweight predictors to drive a 4-cell decision over trust PK, trust CK, trust either, or abstain. Across four LLM backbones, SABER improves end-to-end accuracy and conflict-specific faithfulness over ten inference-time and fine-tuning baselines, with the largest gains on conflict-heavy datasets. Under abstention, SABER's risk-coverage curve Pareto-dominates every prompt-based abstainer, providing a tunable balance between coverage and answer risk. Our code is available at https://github.com/xizhu1022/SABER.

Geometric problem solving, as a typical multimodal reasoning problem, has attracted much attention and made great progress recently, however most of works focus on plane geometry while usually fail in solid geometry due to 3D spatial diagrams and complex reasoning. To bridge this gap, we introduce Hilbert-Geo, the first unified formal language framework for solid geometry, including an extensive predicate library and a dedicated theorem bank. Based on this framework, we propose a Parse2Reason method containing two steps of first parsing then reasoning. In the parsing step, we utilize conditional description language (CDL), a formalized language composed of predicates specifically designed to construct geometric conditions, to represent both problem description (natural text) and solid diagrams (visual image). In the reasoning step, we leverage those formal CDL and the theorem bank to perform relational inference and algebraic computation, generating strictly correct, verifiable, and human-readable reasoning processes. Notably, our proposed Hilbert-Geo is also applicable to plane geometry. To advance geometric reasoning, we curate two expert-annotated dataset SolidFGeo2k and PlaneFGeo3k, which are furnished with geometric formal language annotations, solutions and answers. Extensive experiments show that our proposed method achieves the state-of-the-art (SOTA) performance 77.3% in SolidFGeo2k and 84.1% in MathVerse-Solid (one small subset in MathVerse dedicated to solid geometry), substantially outperforming leading MLLMs, such as Gemini-2.5-pro (54.2% on SolidFGeo2k) and GPT-5 (62.9% on MathVerse-Solid). In addition, our method achieves the SOTA accuracy 80.2% in PlaneFGeo3k, demonstrating the generality of the Hilbert-Geo in geometric reasoning. Our code and datasets are released at https://github.com/PremiLab-Math/Hilbert-Geo.

Agent memory is typically constructed either offline from curated demonstrations or online from post-deployment interactions. However, regardless of how it is built, an agent faces a cold-start gap when first introduced to a new environment without any task-specific experience available. In this paper, we study pre-task memory construction: whether an agent can build procedural memory before observing any target-environment tasks, using only self-generated synthetic practice. Yet, synthetic interaction alone is insufficient, as without controlling what to practice and what to store, synthetic tasks become redundant, infeasible, and ultimately uninformative, and memory further degrades quickly due to unfiltered trajectories. To overcome this, we present Preping, a proposer-guided memory construction framework. At its core is proposer memory, a structured control state that shapes future practice. A Proposer generates synthetic tasks conditioned on this state, a Solver executes them, and a Validator determines which trajectories are eligible for memory insertion while also providing feedback to guide future proposals. Experiments on AppWorld, BFCL v3, and MCP-Universe show that Preping substantially improves over a no-memory baseline and achieves performance competitive with strong playbook-based methods built from offline or online experience, with deployment cost $2.99\times$ lower on AppWorld and $2.23\times$ lower on BFCL v3 than online memory construction. Further analyses reveal that the main benefit does not come from synthetic volume alone, but from proposer-side control over feasibility, redundancy, and coverage, combined with selective memory updates.

Web search queries concern place far more often than existing labelling schemes suggest, yet the landscape of geospatial web search queries - what people ask of place, and how often - remains poorly characterised at scale. We apply dense sentence embeddings, a lightweight SetFit classifier, and density-based clustering to the full MS MARCO corpus of 1.01 million real Bing queries without prior filtering for toponyms or spatial keywords, identifying 181,827 geospatial queries (18.0%), nearly threefold the 6.17% labelled as Location in the original annotations. The resulting taxonomy of 88 query categories reveals that geospatial web search is dominated by transactional and practical lookups: costs and prices alone account for 15.3% of geospatial queries, nearly twice the size of the entire physical geography theme. Much of this activity - costs, opening hours, contact details, weather, travel recommendations - falls outside the scope of what traditional GIS and knowledge graphs are built to serve. The categories vary substantially in the kind of answer they admit, from deterministic lookups answerable from spatial databases or knowledge graphs to evaluative or temporally volatile queries that require generative or real-time systems. We discuss implications for hybrid retrieval architectures and for benchmarks of geographic reasoning in large language models. We openly release the labelled dataset, classifier, and taxonomy.

LLM-as-Judge systems are widely deployed for automated evaluation, yet practitioners lack reliable methods to know when a judge's verdict should be trusted. Token log-probabilities, the standard post-hoc confidence signal, are unavailable for many commercial LLMs and, even when accessible, saturate above 0.999 with structured JSON output. We introduce VERDI (VERification-Decomposed Inference), a method that extracts confidence from the reasoning trace a structured judge already produces, with no additional inference calls. VERDI decomposes each verification-style evaluation into sub-checks and derives three structural signals: Step-Verdict Alignment, Claim-Level Margin, and Evidence Grounding Score. We combine them with Platt-scaled logistic regression. On three public benchmarks, VERDI achieves AUROC 0.72-0.91 on GPT-4.1-mini and 0.66-0.80 on GPT-5.4-mini. On Qwen3.5-4B/9B/27B, where answer-token logprobs are anti-calibrated (higher confidence on errors, AUROC 0.32-0.49), VERDI achieves 0.56-0.70. We additionally validate on a production system with eight rubrics (AUROC 0.73-0.88 on factual rubrics), demonstrate cross-model transfer (AUROC 0.66-0.69), and show that a 33M-parameter NLI (Natural Language Inference) model provides a scalable alternative to regex extraction.

Large Language Models (LLMs) are increasingly deployed in multi-turn dialogue settings where preserving conversational context across turns is essential. A standard serving practice concatenates the full dialogue history at every turn, which reliably maintains coherence but incurs substantial cost in latency, memory, and API expenditure, especially when queries are routed to large proprietary models. Existing approaches often struggle to balance the trade-off between response quality and efficiency. We propose a framework that exploits the early turns of a session to estimate a local response manifold and then adapt a smaller surrogate model to this local region for the remainder of the conversation. Concretely, we learn soft prompts that maximize semantic divergence between the large and surrogate small language models' responses to surface least-aligned local directions, stabilize training with anti-degeneration control, and distill the mined cases into localized LoRA fine-tuning so the surrogate runs without prompts at inference. A simple gate enables a one-time switch with rollback on drift. We further provide a theoretical analysis for key components in SOMA. Extensive experiments show the effectiveness of SOMA. The source code is provided at: https://github.com/LabRAI/SOMA.

We investigate the use of Large Language Models (LLMs) for zero-shot prediction of Ryff Psychological Well-Being (PWB) scores from spontaneous speech. Using a few minutes of voice recordings from 111 participants in the PsyVoiD database, we evaluated 12 instruction-tuned LLMs, including Llama-3 (8B, 70B), Ministral, Mistral, Gemma-2-9B, Gemma-3 (1B, 4B, 27B), Phi-4, DeepSeek (Qwen and Llama), and QwQ-Preview. A domain-informed prompt was developed in collaboration with experts in clinical psychology and linguistics. Results show that LLMs can extract semantically meaningful cues from spontaneous speech, achieving Spearman correlations of up to 0.8 on 80% of the data. Additionally, to enhance explainability, we conducted statistical analyses to characterise prediction variability and systematic biases, alongside keyword-based word cloud analyses to highlight the linguistic features driving the models' predictions.

We study language generation in the limit under a global preference ordering on strings, as introduced by Kleinberg and Wei. As is done in previous work, we aim for breadth, but impose an additional requirement of timeliness: higher-ranked strings should be generated earlier. A string is then only credited if it is generated before a deadline, where its deadline is defined by a function that maps a string's rank in the target language to the time by which it must be produced. This is in keeping with a central consideration in machine learning, where inductive bias favors “simpler” or “more plausible” outputs, all else being equal. We show that timely generation is impossible in a strong sense for eventually consistent generators – the protagonists of most prior related work. Under what is perhaps the mildest natural relaxation of consistency, a hallucination rate that vanishes over time, we show that we can circumvent our impossibility result. In particular, we can achieve optimal density with respect to any superlinear deadline function. We also show this is tight by ruling out timely generation with linear deadlines and vanishing hallucination rate.

Multimodal large language models (MLLMs) have heterogeneous strengths across OCR, chart understanding, spatial reasoning, visual question answering, cost, and latency. Effective MLLM routing therefore requires more than estimating query difficulty: a router must match the multimodal requirements of the current image-question input with the capabilities of each candidate model. We propose LatentRouter, a router that formulates MLLM routing as counterfactual multimodal utility prediction. Given an image-question query, LatentRouter extracts learned multimodal routing capsules, represents each candidate MLLM with a model capability token, and performs latent communication between these states to estimate how each model would perform if selected. A distributional outcome head predicts model-specific counterfactual quality, while a bounded capsule correction refines close decisions without allowing residual signals to dominate the prediction. The resulting utility-based policy supports performance-oriented and performance-cost routing, and handles changing candidate pools through shared per-model scoring with availability masking. Experiments on MMR-Bench and VL-RouterBench show that LatentRouter outperforms fixed-model, feature-level, and learned-router baselines. Additional analyses show that the gains are strongest on multimodal task groups where model choice depends on visual, layout-sensitive, or reasoning-oriented requirements, and that latent communication is the main contributor to the improvement. The code is available at: https://github.com/LabRAI/LatentRouter.

Code generation is typically trained in the primal space of programs: a model produces a candidate solution and receives sparse execution feedback, often a single pass/fail bit. Test-time scaling enriches the inference procedure by sampling multiple candidates and judging among them, but the comparative information this process reveals is discarded after inference. We argue that this information defines a dual judgment space that provides a far richer training signal: the model learns not from an isolated success or failure, but from the relative correctness structure across its own plausible attempts, identifying which succeed, which fail, and what distinguishes them. We introduce DuST (Dual Self-Training), a framework for self-training from the dual judgment space. DuST samples candidate programs from the model's own distribution, labels them through sandbox execution, retains groups containing both successes and failures, and trains the model to rank candidates by execution correctness using GRPO. The objective is purely discriminative: the model is never directly rewarded for generating correct programs. Dual self-training improves both judgment and generation. Across five models spanning two families and three scales (4B to 30B), DuST consistently improves Best-of-4 test-time scaling on LiveCodeBench. For Qwen3-30B-Thinking on LiveCodeBench v6, judgment quality improves by +6.2 NDCG, single-sample pass@1 improves by +3.1, and Best-of-4 accuracy improves by +4.1. The trained model's single rollout matches the base model's Best-of-4 performance. SFT on the same ranking data improves judgment without improving generation, confirming that on-policy RL is the mechanism that transfers dual-space learning back into primal generation.

Capability distillation applies knowledge distillation to selected model capabilities, aiming to compress a large language model (LLM) into a smaller one while preserving the abilities needed for a downstream task. However, most existing methods treat capabilities as independent training targets and overlook how improving one capability can reshape the student's broader capability profile, especially when multiple abilities jointly determine task success. We study capability distillation under a fixed token budget and identify two consistent patterns: distillation induces systematic, budget-dependent cross-capability transfer, and additional budget often brings limited task-relevant gains while sometimes degrading other useful abilities. Building on these insights, we propose ReAD, a Reinforcement-guided cApability Distillation framework that explicitly accounts for capability interdependence. ReAD first infers task-essential capabilities, then generates capability-targeted supervision on the fly, and finally uses an uncertainty-aware contextual bandit to adaptively allocate the distillation budget based on expected utility gains. Extensive experiments show that ReAD improves downstream utility under the same token budget while reducing harmful spillover and wasted distillation effort compared to strong baselines. Our code is publicly available at https://github.com/LabRAI/ReAD.

Autonomous science promises to augment scientific discovery, particularly in complex fields like biomedicine. However, this requires AI systems that can consistently generate novel and diverse solutions to open-ended problems. We evaluate LLMs on the task of open-ended solution generation and quantify their tendency to mode collapse into low-diversity generations. To mitigate this mode collapse, we introduce analogical reasoning (AR) as a new approach to solution generation. AR generates analogies to cross-domain problems based on shared relational structure, then uses those analogies to search for novel solutions. Compared to baselines, AR discovers significantly more diverse generations (improving solution diversity metrics by 90-173%), generates novel solutions over 50% of the time (compared to as little as 1.6% for baselines), and produces high-quality analogies. To validate the real-world feasibility of AR, we implement AR-generated solutions across four biomedical problems, yielding consistent quantitative gains. AR-generated approaches achieve a nearly 13-fold improvement on distributional metrics for perturbation effect prediction, outperform all baselines on AUPRC when predicting cell-cell communication, infer brain region interactions with a high Spearman correlation ($\rho$=0.729) to published methods, and establish state-of-the-art performance on 2 datasets for oligonucleotide property prediction. The novel and diverse solutions produced by AR can be used to augment the search space of existing solution generation methods.

We present Hebatron, a Hebrew-specialized open-weight large language model built on the NVIDIA Nemotron-3 sparse Mixture-of-Experts architecture. Training employs a three-phase easy-to-hard curriculum with continuous anti-forgetting anchoring, followed by supervised fine-tuning on 2 million bilingual Hebrew–English samples. The curriculum ordering alone yields a 3-point aggregate benchmark gain over the reversed configuration. Hebatron achieves a Hebrew reasoning average of 73.8%, outperforming DictaLM-3.0-24B-Thinking (68.9%) and remaining competitive with Gemma-3-27B-IT on GSM8K-HE and Israeli Trivia, while activating only 3B parameters per forward pass across a 30B-parameter model, delivering approximately 9 times higher inference throughput at native context lengths up to 65,536 tokens. To our knowledge, this is the first language-specific adaptation of the Nemotron-3 architecture for any target language, and the first open-weight Hebrew-specialized MoE model with native long-context support. Model weights are released openly to support further research in Hebrew and Semitic-language NLP.

In this paper, we present the RETUYT-INCO participation at the BEA 2026 shared task "Rubric-based Short Answer Scoring for German". Our team participated in track 1 (Unseen answers three-way), track 3 (Unseen answers two-way) and track 4 (Unseen questions two-way). Since these tracks required scoring short student answers using specific rubrics, we looked for ways to handle the changing nature of the task. We created a method called Meta-prompting. In this approach, an LLM creates a custom prompt based on examples from the Train set. This prompt is then used to grade new student answers. Along with this method, we also describe other approaches we used, such as classic machine learning, fine-tuning open-source LLMs, and different prompting techniques. According to the official results, our team placed 6th out of 8 participants in Track 1 with a QWK of 0.729. In Track 3, we secured 4th place out of 9 with a QWK of 0.674, and we also placed 4th out of 8 in Track 4 with a QWK of 0.49.

Computer-use agents (CUAs) rely on visual observations of graphical user interfaces, where each screenshot is encoded into a large number of visual tokens. As interaction trajectories grow, the token cost increases rapidly, limiting the amount of history that can be incorporated under fixed context and compute budgets. This has resulted in no or very limited improvement in the performance when using history unlike other domains. We address this inefficiency by introducing ReVision, which is used to train multimodal language models on trajectories where redundant visual patches are removed using a learned patch selector that compares patch representations across consecutive screenshots while preserving spatial structure required by the model. Across three benchmarks, OSWorld, WebTailBench, and AgentNetBench, when processing trajectories with 5 history screenshots using Qwen2.5-VL-7B, ReVision reduces token usage by 46% on average while improving success rate by 3% over the no drop baseline. This establishes a clear efficiency gain, enabling agents to process longer trajectories with fewer tokens. With this improved efficiency, we revisit the role of history in CUAs and find that performance continues to improve as more past observations are incorporated when redundancy is removed.

Instructions trigger a production-centered mechanism in language models. Through a cognitively inspired lens that separates language processing and production, we reveal this mechanism as an asymmetry between the two stages by probing task-specific information layer-wise across five binary judgment tasks. Specifically, we measure how instruction tokens shape information both when sample tokens, the input under evaluation, are processed and when output tokens are produced. Across prompting variations, task-specific information in sample tokens remains largely stable and correlates only weakly with behavior, whereas the same information in output tokens varies substantially and correlates strongly with behavior. Attention-based interventions confirm this pattern causally: blocking instruction flow to all subsequent tokens reduces both behavior and information in output tokens, whereas blocking it only to sample tokens has minimal effect on either. The asymmetry generalizes across model families and tasks, and becomes sharper with model scale and instruction-tuning, both of which disproportionately affect the production stage. Our findings suggest that understanding model capabilities requires jointly assessing internals and behavior, while decomposing the internal perspective by token position to distinguish the processing of input tokens from the production of output tokens.

Large language models (LLMs) trained on web-scale corpora can memorize sensitive training data, posing significant privacy risks. Differential privacy (DP) has emerged as a principled framework that limits the influence of individual data points during training, yet the relationship between differential privacy and social bias in LLMs remains poorly understood. To investigate this, we present a systematic evaluation of social bias in a pretrained LLM trained with DP-SGD, comparing a DP model against non-DP baselines across four complementary paradigms: sentence scoring, text completion, tabular classification, and question answering. We find that DP reduces bias in sentence scoring tasks, where bias is measured through controlled likelihood comparisons, yet this improvement does not generalize across all tasks. Our results reveal a discrepancy between logit-level bias and output-level bias. Moreover, decreasing memorization does not necessarily reduce unfairness, underscoring the importance of multi-paradigm evaluation when assessing fairness in LLMs.

Existing multi-model and tool-augmented systems communicate by generating text, serializing every exchange through the output vocabulary. Can two pretrained language models instead coordinate through a continuous, concurrent channel? The Bicameral Model couples two frozen language models through a trainable neural interface on their intermediate hidden states. At every generation step, both models run in lockstep: a primary model drives the task while an auxiliary model operates tools, solves constraints, or executes code, with both conditioning on each other's activations through a translation network and a learned suppression gate ($\sim$1% of combined parameters). The gate learns a selective communication protocol from task loss alone, without a prescribed format. We demonstrate the mechanism across three tool backends. On arithmetic, coupling two 0.5B models with a calculator raises accuracy from 36% to 96%. On logic grid puzzles, coupling two 0.6B models with a Z3 solver achieves $1.7\times$ the unaugmented baseline on ZebraLogic. On mathematical reasoning, coupling with a Python sandbox enables the auxiliary to generate problem-specific code from hidden-state signals alone, without ever seeing the problem text.

We decompose an evolutionary mixture-of-LoRA system on a from-scratch ~150M-parameter widened-D substrate (D=1536, V=32000; D/V approx 0.048; the "widened-1536" substrate) into three factors – a router rewrite (parallel sigmoid gate with learnable per-adapter floor and bounded temperature anneal, fed post-stack hidden states rather than token-embedding means), a per-domain leave-one-out evaluation scope, and a lifecycle of death plus alpha-blend inheritance plus SVD mutation plus slot reallocation – and report a 5-of-8 partial 2^3 factorial run at n=3 seeds and 25000 adaptation steps per cell. The attribution chain is sharp on this substrate: the router rewrite carries the entire +0.0426 nat balanced log-PPL improvement (Delta = log PPL_ref - log PPL_test, positive = improvement; t=12.86, p=0.006) attributed to "the full evolutionary system vs the static B3 baseline"; the headline full-system-vs-B3 balanced contrast itself is +0.015 nats, t=1.94, p=0.19 at n=3 and does not clear alpha=0.05. The per-domain evaluation scope is null at seed-resolution, and the lifecycle is a net drag of approx -0.028 nats (t=-4.46,p=0.047 in the primary chain). An auxiliary alpha=0 inheritance counterfactual at n=3 seeds is sign-inconsistent at the headline metric and underpowered for either an equivalence or load-bearing conclusion (corrected from an earlier arithmetic-mean aggregator that erroneously cleared inheritance; see Appendix B.11). A base-perturbation probe directionally refutes a "genomic-context" reframe of the lifecycle role. A controllable synthetic sandbox locates a substrate-conditional regime boundary: evolutionary search on the routing channel is load-bearing only when adapters are pre-aligned to the task; in every other regime tested it underperforms, ties, or actively degrades the gradient solution.

Reasoning benchmarks measure clinical performance on clean inputs. We evaluate the step before reasoning: retrieval over real EHR notes, where negation, temporality, and family-versus-patient attribution can flip a correct answer to a wrong one. EpiKG carries an assertion label and a temporality tag with every fact in a patient knowledge graph, then routes retrieval by question intent. ClinicalBench is a 400-question test over 43 MIMIC-IV patients across 9 assertion-sensitive categories. A 7-condition ablation tests each piece of EpiKG across six LLMs (Claude Opus 4.6, GPT-OSS 20B, MedGemma 27B, Gemma 4 31B, MedGemma 1.5 4B, Qwen 3.5 35B). Three physicians blindly adjudicated 100 paired items. The author-blind primary endpoint, leave-author-out paired exact McNemar on 50 unanimous-strict items rated by two external physicians, yields +22.0 percentage points (95 percent Newcombe CI [+5.1, +31.5], p=0.0192). The architectural novelty, intent-aware KG-RAG over a Contriever dense-RAG baseline (C2b to C4g_kw on the change-excluded n=362 endpoint), is +8.84 percentage points (paired McNemar p=1.79e-3); +12.43 percentage points under oracle intent. Sensitivities agree directionally: three-rater physician majority +24.0 percentage points (subject to single-author circularity); deterministic keyword reproducibility proxy +39.5 percentage points. Across the six models, the gain shrinks as the LLM-alone baseline rises (beta=-1.123, r=-0.921, p=0.009). With n=6 this looks more like regression to the mean than encoding substituting for model size. Physician adjudication identified 56 percent of auto-generated reference answers as defective, a methodological finding indicating that NLP-pipeline clinical-QA benchmarks require physician adjudication to be usable. ClinicalBench, the frozen evaluator, three-rater adjudication data, and the EpiKG output stack are publicly released.