Papers

Multi-agent systems (MAS) are utilized in many contexts and many professions. Those MAS rely on inter-agent communication, usually implemented by clear-text message passing. We hypothesize that Large Language Models may have a world model at their disposal that exceeds expressibility in text when complex concepts need to be communicated. Our aim is to approach a proof of this hypothesis with structured experiments. In this work, we show that LLM agents communicating via text lose information, which we quantify via Sparse Autoencoder (SAE) feature analysis. We construct three communication channels and measure concept-discriminating information in each. We first show that the SAE-sparse channel retains a 99.4% probe accuracy at 28-fold compression over the dense-latent channel vs 80.4% for the text channel. We then proceed to examine the same for cross-architecture communication by using sparse latent space alignment. We find for Procrustes alignment a 92% top-1 retrieval between Llama and Mistral. Using a text round-trip, we perform feature survival analysis to find that text serialization destroys 88% of SAE features, replacing them with a different feature set. We attribute the loss to identity replacement, not attenuation. By our analysis, we were able to attribute a 3-10pp performance penalty to the linear Procrustes alignment, improving with nonlinear alignment methods. In a task-level evaluation we find that the latent channel matches the text channel on cross-lingual concept tasks but never exceeds it. Text augmentation with latent features provides no benefit, leading us to negative conclusions for the initial hypothesis: lost features mostly or completely encode surface form, not task-relevant semantics. To pinpoint the practical advantage of latent communication over a text channel, deeper tasks eliciting complex concepts and an corresponding analysis framework are needed.

With the rapid growth of digital data, real-world applications increasingly involve hierarchical information that combines static attributes with dynamic records. Modeling such heterogeneous data in a unified and generalizable manner remains challenging. Existing approaches often rely on extensive manual design, are tightly coupled to specific data schemas, and typically process static and dynamic attributes in isolation, thereby overlooking their implicit interactions. We propose UniSAGE, a unified framework for modeling data with both static and dynamic attributes. UniSAGE constructs a global attribute graph that represents hierarchical and temporal relationships in a unified structure. To ensure representational consistency, it introduces two orthogonal parameter subspaces that jointly support static aggregation and dynamic reasoning within a shared semantic space. Building on these unified representations, UniSAGE further enables task-specific interaction between static and dynamic attributes via a lightweight hyper-structure mechanism. UniSAGE is fully automated, robust to evolving data schemas, and capable of capturing complex cross-attribute dependencies. Extensive experiments on multiple public benchmarks and a real-world financial behavior dataset demonstrate that UniSAGE consistently outperforms existing methods, achieving performance improvements of over 10% on several tasks.

Large language models (LLMs) are increasingly used for decision support in healthcare, but clinical evidence is often incomplete or evolving. When the available information is insufficient to support a reliable answer, models should request clarification or abstain rather than provide unsupported responses. Existing medical benchmarks, however, typically assume that complete information is available upfront. We introduce Safe-Psych, a sequential benchmark for evaluating how LLMs handle evolving diagnostic uncertainty in clinical psychiatry. Safe-Psych contains over 1,000 real-world psychiatric clinical notes segmented to simulate incremental evidence disclosure, with psychiatrist-derived action labels at each stage: DIAGNOSE, CLARIFY, or ABSTAIN. We evaluate multiple state-of-the-art LLMs in full-information and sequential settings. Our findings show that capability does not ensure calibration: even strong models struggle under incomplete clinical information, with under-abstention exceeding 60% for most models and safety-aware prompting reducing premature commitment only by shifting errors toward excessive abstention. In sequential evaluation, models frequently diagnose before sufficient evidence is available and rarely seek clarification unless explicitly prompted; these premature diagnoses are less accurate than on-time diagnoses. Overall, Safe-Psych reveals a limitation across the evaluated models: recognizing when clinical evidence is incomplete and additional information is needed. We release Safe-Psych to support research on improving LLM safety in healthcare.

The deployment of large language models (LLMs) in specialized domains like medical diagnostics and financial advisory necessitates evaluating capabilities beyond general knowledge. Traditional question-answering benchmarks often fail to capture the nuanced contextual grounding, user awareness, and domain understanding these fields require. To address this, we introduce CANDI-QA (Contextual Alignment for Niche Domains Question Answering), a novel dataset evaluating LLMs on delivering accurate, context-sensitive, and user-aligned answers in specialized settings. CANDI-QA features expert-curated question-answer pairs structured into two categories: (1) Information Assistance Questions, which are direct, factual queries requiring precise extraction, and (2) Applied Inference Questions, which are multi-hop reasoning tasks needing situational inference to generate actionable insights. We evaluate over ten diverse language models, from compact open-source to state-of-the-art proprietary systems. As a robust baseline, we present MTSS-Net, a lightweight neuro-symbolic framework combining neural retrieval with rule-based reasoning. Our findings highlight the profound challenges of achieving contextual alignment in niche domains, revealing the limitations of current LLMs without enhanced contextual or symbolic integration. Ultimately, CANDI-QA serves as a critical benchmark for advancing research in context-aware language models, stimulating the development of robust, trustworthy AI for high-stakes domains.

Students at all levels of higher education face a significant barrier in the form of information overload, which often paralyzes the initial stages of the research process and suppresses motivation. In response, this article introduces a pedagogical framework that leverages AI SciBrief, a platform powered by a Large Language Model (LLM) designed to automatically generate digests of scientific trends. We describe how this multidisciplinary tool - with initial coverage in finance, medicine, and education - can be integrated into the curriculum to overcome this "entry barrier." The framework provides concrete methodologies for utilizing these digests to facilitate topic selection for term papers, accelerate literature reviews for dissertations, and enable postgraduate students to continuously monitor emerging trends. We conclude that AI SciBrief functions as a "gateway to research" effectively reducing students' cognitive load and empowering them to transition more rapidly from information searching to knowledge creation.

The automated extraction of data from scientific charts is a critical task for large-scale literature analysis. While multimodal Large Language Models (LLMs) show promise, their accuracy on non-standardized charts remains a challenge. This raises a key research question: what is the most effective strategy to improve model performance (high-level semantic priming) or low-level spatial priming? This paper presents a comparative investigation into these two distinct strategies. We describe our exploratory experiments with semantic methods, such as a two-stage metadata-first framework and Chain-of-Thought, which failed to produce a statistically significant improvement. In contrast, we present a simple but highly effective spatial priming method: overlaying a coordinate grid onto the chart image before analysis. Our quantitative experiment on a synthetic dataset demonstrates that this grid-based approach provides a statistically significant reduction in data extraction error (SMAPE reduced from 25.5% to 19.5%, p < 0.05) compared to a baseline. We conclude that for the current generation of multimodal models, providing explicit spatial context is a more effective and reliable strategy than high-level semantic guidance for this class of tasks.

There is an increasing number of algorithmic computations in theoretical physics. These, while conceptually simple, can nevertheless be time-consuming and contain subtleties that should not be overlooked. Given the recent improvement of Large Language Models (LLM), it is natural to investigate whether LLMs equipped with a computer algebra system (CAS) runtime and sufficiently informative context can reliably carry out these algorithmic tasks. In this work, we interface Claude with Maple, and apply this framework to cosmological perturbations in modified theories of gravity. We demonstrate the current capabilities of this approach, the typical failures, and how the same can be improved. We find that a frontier LLM supplied with worked examples is able to solve most test problems.

Image-Text-Matching (ITM) is one of the defacto methods of learning generalized representations from a large corpus in Vision and Language (VL). However, due to the weak association between the web-collected image-text pairs, models fail to show a fine-grained understanding of the combined semantics of these modalities. To address this issue we propose Hard Negative Captions (HNC): an automatically created dataset containing foiled hard negative captions for ITM training towards achieving fine-grained cross-modal comprehension in VL. Additionally, we provide a challenging manually-created test set for benchmarking models on a fine-grained cross-modal mismatch task with varying levels of compositional complexity. Our results show the effectiveness of training on HNC by improving the models' zero-shot capabilities in detecting mismatches on diagnostic tasks and performing robustly under noisy visual input scenarios. Also, we demonstrate that HNC models yield a comparable or better initialization for fine-tuning

Statistical watermarking is a common approach for verifying whether text was written by a language model. Most existing schemes assume autoregressive generation, where tokens are produced left to right and contextual hashing is well defined. Diffusion language models generate text by denoising tokens in arbitrary order, so these schemes cannot be applied directly. A recent watermark by Gloaguen et al. addresses this gap for LLaDA 8B Instruct and reports true positive detection above 99%. This paper studies what happens when watermarked text is rewritten not once but several times. Using the same watermark configuration, 1,605 watermarked completions of about 300 tokens each are produced across five WaterBench domains. Each completion is rewritten by four open weight language models, from 1.5B to 8B parameters, none of which know the watermark key. Five rewrite styles are tested: paraphrase, humanize, simplify, academic, and summarize expand. Each style is chained for up to five hops, producing 160,500 rewritten texts in total. The watermark is detected on 87.9% of the original outputs at the standard significance threshold. After a single rewrite, detection falls to between 14% and 41% depending on the rewriter and style. After five chained rewrites, detection falls to 4.86%, meaning 94.76% of the originally detected texts are no longer flagged. After three rewrites, the detector score has dropped 86% of the way from its watermarked baseline toward the null distribution. Repeated rewriting is therefore a much stronger attack than a single rewrite, and the result holds across all four rewriters tested.

LLMs can solve program synthesis tasks but remain inefficient and unreliable on hard instances requiring large combinatorial search. Given a small set of reasoning traces, we use coding agents to compile them into reusable symbolic program synthesizers over constrained DSLs. The resulting solvers require no LLM calls at test time and are strong standalone systems: symbolic solver ensembles reach 91.3% accuracy on PBEBench-Lite and 84.7% on PBEBench-Hard, outperforming LLMs with test-time scaling for the latter by +16.3 percentage points at zero LLM inference cost. They also complement LLM search, improving PBEBench-Hard accuracy from 68.4% to 85.8% while reducing reported token usage by 78%, and raising SLR-Bench hard-tier accuracy from 34.4% to 58.0% in a neuro-symbolic hybrid setting. Compared to directly using coding agents as per-instance solvers, induced solvers are substantially more Pareto-efficient, amortizing a small one-time construction cost over many zero-token executions. Finally, most solvers transfer zero-shot to a real historical linguistics task - predicting sound changes in natural language data - reaching 80.1% accuracy under ensembling and recovering some plausible linguistic rules. Together, these results show that reasoning traces can be compiled into reusable symbolic solvers that solve many tasks directly, complement LLM inference on hard cases, and provide a scalable route to domain-general solver induction. We release code and data for reproducibility.

Large language models (LLMs) are rapidly being adopted across various domains. However, their adoption in banking industry faces resistance due to demands for high accuracy, regulatory compliance, and the need for verifiable and grounded responses. We present a unified, data-efficient framework for training grounded domain-specific LLMs that optimizes answer quality, citation grounding, and calibrated refusal under real-world deployment constraints. First, we describe a data generation pipeline that combines LLM-as-a-Judge filtering, citation annotation, and curriculum learning with only 143M tokens. The resulting 12B model achieves high answer quality outperforming GPT-4.1 on citation grounding, with a modest citation tradeoff versus the untuned base. Second, we propose a calibrated refusal mechanism: training on 22% unanswerable examples yield a 12% "I don't know" rate, substantially improving over the base model's unsafe 4.3% rate while avoiding GPT-4.1's over-refusal (20.2%). Third, we present an end-to-end methodology spanning from data curation to quantized serving. The system is deployed at 40+ financial institutions, achieving a 7.1 percentage point improvement in query resolution (p < 0.001). Additionally, the model delivers 3-5x faster responses at 20-50x lower cost compared to GPT-4.1.

Knowledge graphs automatically constructed from text are increasingly used in real-world applications. However, their inherent noise, fragmentation, and semantic inconsistencies significantly affect the performance of Graph Neural Networks (GNNs) on downstream tasks. Assessing their performance and robustness remains difficult, as it is often unclear whether observed results stem from the learning model or from the quality of the constructed graph itself. In this work, we introduce a dual-purpose benchmark designed to jointly evaluate (i) the performance of GNNs on noisy, text-derived graphs and (ii) the effectiveness of graph construction methods on a downstream task. The benchmark is built in the biomedical domain from a single textual corpus and includes two automatically constructed graphs generated using different extraction methods, alongside a high-quality reference graph curated by experts that serves as an upper performance bound. This design enables controlled comparison of construction methods and systematic evaluation of GNN robustness through semi-supervised node classification. We further provide a standardized, reproducible, and extensible evaluation framework, facilitating the integration of new graph extraction methods and learning models.

LLM watermarks must be detectable without compromising text quality, yet most existing schemes bias the next-token distribution and pay for detection with measurable quality loss. We present SLAM (Structural Linguistic Activation Marking), a novel white-box watermarking scheme that sidesteps this cost by writing the mark into structural geometry rather than token frequencies: sparse autoencoders identify residual-stream directions encoding linguistic structure (e.g., voice, tense, clause order), and we causally steer those directions at generation time, leaving lexical sampling and semantics unconstrained. On Gemma-2 2B and 9B, SLAM achieves 100% detection accuracy with a quality cost of only 1-2 reward points - compared to 7.5-11.5 for KGW, EWD, and Unigram - with naturalness and diversity preserved at near-unwatermarked levels across both models. The trade-off is a complementary robustness profile: SLAM resists word-level edits but is vulnerable to paraphrase that restructures syntax (at a quality cost), the converse of token-distribution methods.

Financial document question answering (QA) demands complex multi-step numerical reasoning over heterogeneous evidence–structured tables, textual narratives, and footnotes–scattered across corporate filings. Existing retrieval-augmented generation (RAG) approaches adopt a single-pass retrieve-then-generate paradigm that struggles with the compositional reasoning chains prevalent in financial analysis. We propose FinAgent-RAG, an agentic RAG framework that orchestrates iterative retrieval-reasoning loops with self-verification, specifically engineered for the precision requirements of financial numerical reasoning. The framework integrates three domain-specific innovations: (1) a Contrastive Financial Retriever trained with hard negative mining to distinguish semantically similar but numerically distinct financial passages, (2) a Program-of-Thought reasoning module that generates executable Python code for precise arithmetic rather than relying on error-prone LLM-based mental computation, and (3) an Adaptive Strategy Router that dynamically allocates computational resources based on question complexity, reducing API costs by 41.3% on FinQA while preserving accuracy. Extensive experiments on three benchmark datasets–FinQA, ConvFinQA, and TAT-QA–demonstrate that FinAgent-RAG achieves 76.81%, 78.46%, and 74.96% execution accuracy respectively, outperforming the strongest baseline by 5.62–9.32 percentage points. Ablation studies, cross-backbone evaluation with four LLMs, and deployment cost analysis confirm the framework's robustness and practical viability for financial institutions.

Large-scale datasets are widely used to perform summarization tasks, but they may not include queries alongside documents and summaries. In the search for suitable datasets for Query-Focused Summarization (QFS), we identify two research questions: Is it possible to automatically generate evidence-based query keywords from query-free datasets? Does evidence-based query generation support the QFS task? This paper proposes an evidence-based model to generate queries from query-free datasets. To evaluate our model intrinsically, we compare the similarity between the original queries and the system-generated queries of two QFS datasets. We also perform summarization tasks using different pre-trained models, as well as a state-of-the-art (SOTA) QFS model, to measure the extrinsic performance of our query generation approach. Experimental results indicate that summaries generated using evidence-based queries achieve competitive ROUGE scores compared to those generated from the original queries.

Large language models increasingly operate in interactive settings where solving a task requires multiple rounds of information exchange with a user. However, most current systems treat dialogue reactively and lack a principled mechanism to reason about what information is missing. We propose BALAR (Bayesian Agentic Loop for Active Reasoning), a task-agnostic outer-loop algorithm that requires no fine-tuning and enables multi-turn interaction between an LLM agent and a user. BALAR maintains a structured belief over latent states, selects clarifying questions by maximizing expected mutual information, and dynamically expands its state representation when the current one proves insufficient. We evaluate BALAR on three diverse benchmarks: AR-Bench-DC (detective cases), AR-Bench-SP (thinking puzzles), and iCraft-MD (clinical diagnosis). BALAR outperforms all baselines across the three benchmarks, with 14.6% higher accuracy on AR-Bench-DC, 38.5% on AR-Bench-SP, and 30.5% on iCraft-MD. We further study whether BALAR can serve as a teacher for a questioning policy through supervised fine-tuning (SFT), direct preference optimization (DPO), and dense-reward reinforcement learning (RL). Across 18 iCraft-MD replications, distilling BALAR into a Llama-8B yields relative gains in frozen-Qwen final-answer accuracy of 8.1% with SFT, 10.9% with DPO, and 12.1% with RL over the untuned policy.

We present ZAYA1-8B, a reasoning-focused mixture-of-experts (MoE) model with 700M active and 8B total parameters, built on Zyphra's MoE++ architecture. ZAYA1-8B's core pretraining, midtraining, and supervised fine-tuning (SFT) were performed on a full-stack AMD compute, networking, and software platform. With under 1B active parameters, ZAYA1-8B matches or exceeds DeepSeek-R1-0528 on several challenging mathematics and coding benchmarks, and remains competitive with substantially larger open-weight reasoning models. ZAYA1-8B was trained from scratch for reasoning, with reasoning data included from pretraining onward using an answer-preserving trimming scheme. Post-training uses a four-stage RL cascade: reasoning warmup on math and puzzles; a 400-task RLVE-Gym curriculum; math and code RL with test-time compute traces and synthetic code environments built from competitive-programming references; and behavioral RL for chat and instruction following. We also introduce Markovian RSA, a test-time compute method that recursively aggregates parallel reasoning traces while carrying forward only bounded-length reasoning tails between rounds. In TTC evaluation, Markovian RSA raises ZAYA1-8B to 91.9% on AIME'25 and 89.6% on HMMT'25 while carrying forward only a 4K-token tail, narrowing the gap to much larger reasoning models including Gemini-2.5 Pro, DeepSeek-V3.2, and GPT-5-High.

This intervention study investigates the use of counterarguments in writing for critical thinking by students in the context of Generative AI (GenAI). This is especially as risks of cheating and cognitive offloading exist with the use of GenAI. We presented 36 students in a particular university course with 4 carefully selected thesis statements (from a set of popular debates) to write about anyone of them. We used six established rubrics (focus, logic, content, style, correctness and reference) to conduct three human assessments (two student peer-reviews and one experienced teacher) per writeup on a 5-point Likert scale for all the qualified samples (n) of 35 submissions (after disqualifying one for irregularity). Using the same rubrics and guidelines, we also assessed the submissions using six frontier LLMs as judges. Our mixed-method design included qualitative open-ended feedback per assessment and quantitative methods. The results reveal that (1) the students' self-written counterarguments to AI-generated content contains logic, among other things, which is a key component of critical thinking, and (2) GenAI can be successfully used at scale to assess students' written work, based on clear rubrics, and these assessments generally align with human assessments as shown with Gwets AC2 inter-rater reliability values of 0.33 for all the models except one.

Code translation is one of the core capabilities of LLMs. However, evaluating the correctness of translations remains difficult, as commonly used metrics such as BLEU measure only syntactic similarity, disregarding program semantics. We propose a novel evaluation methodology for code translation tasks, emphasizing semantic equivalence over surface-level string similarity. Our approach applies established compiler testing methodology to a new domain, allowing the assessment of an LLM fine-tuned for binary lifting tasks (i.e. decompiling binaries to higher-level representations). We introduce a semantic correctness score, defined as the proportion of translations that produce correct execution outcomes, and demonstrate its application by evaluating LLM-based and heuristic decompilers. Our findings show that LLM-based approaches significantly outperform heuristic ones, while BLEU scores show negligible correlation with semantic correctness (r = -0.127 to 0.354), demonstrating that syntactic metrics fail to predict functional accuracy.

AI-assisted resume tailoring systems commonly operate on a single uploaded resume, which limits their ability to recover relevant experience omitted from the current draft and makes it difficult for users to distinguish grounded edits from model-generated suggestions. This paper presents Resume Tailor, an agentic resume-tailoring system that maintains a longitudinal career vault in a vector database and uses multi-source retrieval-augmented generation (RAG) to assemble job-specific resume content from historical resumes and structured career records. The system is implemented as a 12-node LangGraph pipeline with typed state management, hybrid semantic-lexical confidence scoring, provenance-aware fallback generation, anti-hallucination guardrails, and a conditional review loop. We report a pilot evaluation on nine job descriptions (JDs) across software engineering, data analytics, and business analysis roles using a single candidate's career history. For six JDs where the candidate held at least one prior role in the same occupational category, enabling the career vault improved Applicant Tracking System (ATS)-style fit scores by an average of 7.8 points. For two JDs requiring domain-specific expertise absent from the vault, scores decreased by an average of 8.0 points. One partially overlapping role showed a modest gain of 2 points. These results suggest that longitudinal retrieval can improve resume tailoring when relevant prior experience exists, while also highlighting the need for confidence-gated retrieval when domain overlap is weak.

Grammaticality and likelihood are distinct notions in human language. Pretrained language models (LMs), which are probabilistic models of language fitted to maximize corpus likelihood, generate grammatically well-formed text and discriminate well between grammatical and ungrammatical sentences in tightly controlled minimal pairs. However, their string probabilities do not sharply discriminate between grammatical and ungrammatical sentences overall. But do LMs implicitly acquire a grammaticality distinction distinct from string probability? We explore this question through studying internal representations of LMs, by training a linear probe on a dataset of grammatical and (synthetic) ungrammatical sentences obtained by applying perturbations to a naturalistic text corpus. We find that this simple grammaticality probe generalizes to human-curated grammaticality judgment benchmarks and outperforms LM probability-based grammaticality judgments. When applied to semantic plausibility benchmarks, in which both members of a minimal pair are grammatical and differ in only plausibility, the probe however performs worse than string probability. The English-trained probe also exhibits nontrivial cross-lingual generalization, outperforming string probabilities on grammaticality benchmarks in numerous other languages. Additionally, probe scores correlate only weakly with string probabilities. These results collectively suggest that LMs acquire to some extent an implicit grammaticality distinction within their hidden layers.

Background: Existing MRI LLM benchmarks rely mainly on review-book multiple-choice questions, where top proprietary models already score highly, limiting discrimination. No systematic benchmark has evaluated vendor-specific scanner operational knowledge central to research MRI practice. Purpose: We developed MRI-Eval, a tiered benchmark for relative model comparison on MRI physics and GE scanner operations knowledge using primary multiple-choice questions (MCQ), with stem-only and primed diagnostic conditions as complementary analyses. Methods: MRI-Eval includes 1365 scored items across nine categories and three difficulty tiers from textbooks, GE scanner manuals, programming course materials, and expert-generated questions. Five model families were evaluated (GPT-5.4, Claude Opus 4.6, Claude Sonnet 4.6, Gemini 2.5 Pro, Llama 3.3 70B). MCQ was primary; stem-only removed options and used an independent LLM judge; primed stem-only tested responses to incorrect user claims. Results: Overall MCQ accuracy was 93.2% to 97.1%. GE scanner operations was the lowest category for every model (88.2% to 94.6%). In stem-only, frontier-model accuracy fell to 58.4% to 61.1%, and Llama 3.3 70B fell to 37.1%; GE scanner operations stem-only accuracy was 13.8% to 29.8%. Conclusion: High MCQ performance can mask weak free-text recall, especially for vendor-specific operational knowledge. MRI-Eval is most informative as a relative comparison benchmark rather than an absolute competency measure and supports caution in using raw LLM outputs for GE-specific protocol guidance.

Self-consistency detects hallucinations by generating multiple sampled answers to a question and measuring agreement, but this requires repeated decoding and can be sensitive to lexical variation. Semantic self-consistency improves this by clustering sampled answers by meaning using natural language inference, but it adds both sampling cost and external inference overhead. We show that first-token confidence, phi_first, computed from the normalized entropy of the top-K logits at the first content-bearing answer token of a single greedy decode, matches or modestly exceeds semantic self-consistency on closed-book short-answer factual question answering. Across three 7-8B instruction-tuned models and two benchmarks, phi_first achieves a mean AUROC of 0.820, compared with 0.793 for semantic agreement and 0.791 for standard surface-form self-consistency. A subsumption test shows that phi_first is moderately to strongly correlated with semantic agreement, and combining the two signals yields only a small AUROC improvement over phi_first alone. These results suggest that much of the uncertainty information captured by multi-sample agreement is already available in the model's initial token distribution. We argue that phi_first should be reported as a default low-cost baseline before invoking sampling-based uncertainty estimation.

We present our system for SemEval-2026 Task 9: Multilingual Polarization Detection, a binary classification task spanning 22 languages. Our approach fine-tunes separate Gemma~3 models (12B and 27B parameters) per language using Low-Rank Adaptation (LoRA), augmented with synthetic data generated by a large language model (LLM). We employ three synthetic data strategies (direct generation, paraphrasing, and contrastive pair creation) using GPT-4o-mini, with a multi-stage quality filtering pipeline including embedding-based deduplication. We find that per-language threshold tuning on the development set yields 2 to 4% F1 improvements without retraining. We also use weighted ensembles of 12B and 27B model predictions with per-language strategy selection. Our final system achieves a mean macro-F1 of 0.811 across all 22 languages, ranking 2nd overall of the participating teams, with 1st place finishes in 3 languages and top-3 in 8 languages. We also find that alternative architectures (XLM-RoBERTa, Qwen3) that showed strong development set performance suffered 30 to 50% F1 drops on the test set, highlighting the importance of generalization.

Automated mental health prediction using textual data has shown promising results with deep learning and large language models. However, deploying these models in high-stakes real-world settings remains challenging, as existing approaches largely rely on semantic representations and often produce overconfident predictions under ambiguous, noisy, or shifted data. Moreover, most methods lack reliable uncertainty estimation, undermining trust in risk-sensitive mental health applications. To address these limitations, we formulate the task as a multi-view learning problem that integrates semantic information from encoder-only models with higher-level reasoning information from decoder-only models, where reasoning-aware representations and uncertainty modeling are obtained in a trustworthy manner. To ensure reliable fusion, we adopt an evidential learning framework based on Subjective Logic to explicitly model uncertainty and introduce an evidential fusion strategy that balances complementary views while discounting unreliable evidence. Benchmarking on three real-world datasets, Dreaddit, SDCNL, and DepSeverity, reports accuracies of 0.835, 0.731, and 0.751, respectively, demonstrating its potential for reliable mental health prediction. Additional experiments on robustness to noise and case studies for interpretability confirm that our proposed framework not only improves predictive performance but also provides trustworthy uncertainty estimates and human-understandable reasoning signals, making it suitable for risk-sensitive applications in mental health assessment.

We introduce the Concept Field of a text corpus: a local drift field with pointwise uncertainty, estimated in sentence-embedding space from the deltas between consecutive sentences. Given a candidate sentence transition, we score its agreement with the field by $\zeta$, the mean absolute z-distance between the observed delta and the field's local statistical estimate, giving a quantitative measure of groundedness. The score is model agnostic and corpus-attributable, providing traces to corpus semantics either supporting or refuting groundedness. We drive the computation with the introduction of a Vector Sequence Database (VSDB) that stores embeddings together with sequence-position and next-delta metadata. We demonstrate this approach on two large-scale settings: groundedness detection over the U.S. Code of Federal Regulations, and ungroundedness (novelty) detection over Project Gutenberg, where we show Concept Fields achieve strong selective classification performance under a grounded / ungrounded / unsure triage policy. Unlike retrieval-centric baselines, the resulting coverage-risk behavior shows similar empirical coverage-performance trends in the two evaluated domains. Concept Fields provide a fast, lightweight, and interpretable signal for groundedness, complementary to LLM-as-judge and white-box detectors.

LLMs are trained once, then deployed into a world that never stops changing. External memory compensates for this, but most systems manage it explicitly rather than letting it adapt on its own. Biological memory works differently: coupled multi-timescale dynamics make new associations immediately usable, strengthen what repetition confirms, and let the rest fade. We argue that external memory should follow a similar principle. In Memini, this view takes the form of an associative memory that organizes knowledge as a directed graph. Each edge carries two coupled internal variables, one fast and one slow, following the Benna-Fusi model of synaptic consolidation. From this coupling, episodic sensitivity, gradual consolidation, and selective forgetting are expected to emerge as facets of a single mechanism, reframing external memory as a learning substrate that reorganizes through its own dynamics. This workshop article describes an early-stage conceptual design without experimental evaluation.

We present an automated, contrastive evaluation pipeline for auditing the behavioral impact of interventions on large language models. Given a base model $M_1$ and an intervention model $M_2$, our method compares their free-form, multi-token generations across aligned prompt contexts and produces human-readable, statistically validated natural-language hypotheses describing how the models differ, along with recurring themes that summarize patterns across validated hypotheses. We evaluate the approach in synthetic setting by injecting known behavioral changes and showing that the pipeline reliably recovers them. We then apply it to three real-world interventions, reasoning distillation, knowledge editing and unlearning, demonstrating that the method surfaces both intended and unexpected behavioral shifts, distinguishes large from subtle interventions, and does not hallucinate differences when effects are absent or misaligned with the prompt bank. Overall, the pipeline provides a statistically grounded and interpretable tool for post-hoc auditing of intervention-induced changes in model behavior.

Human psychometric questionnaires are increasingly used to characterize large language models (LLMs), but it remains unclear what exactly their responses reflect. To answer this question we analyze responses from 50 LLMs across 45 questionnaires spanning multiple psychological domains using psychometric methodology. We found a dominant principal component explaining 47.1% of variance across standardized, response-derived questionnaire-level primary scores, revealing substantial shared structure across instruments. We investigate its meaning through semantic analysis, respondent framing, and human annotations. Semantic analysis suggests that the component partly reflects experiential self-attribution: how readily models endorse statements describing subjective states and bodily experiences. A model score emphasizing items whose disagreement decreases under human simulation correlates strongly with the component ($r=0.864$). Human ratings of 289 matched items show that inner-life content is positively associated, and social desirability negatively associated, with item-component correlations after mutual adjustment and controls for other semantic properties. Together, six rated properties explain 22.1% of variance in Fisher-transformed item-component correlations, supporting a partial interpretation involving both experiential self-attribution and socially desirable responding. These findings suggest that LLM psychometric profiles share response tendencies that require interpretation beyond the human traits their questionnaires were designed to measure.

We identify and prove a fundamental trade-off governing long-sequence models: no model can simultaneously achieve (i) per-step computation independent of sequence length (Efficiency), (ii) state size independent of sequence length (Compactness), and (iii) the ability to recall a number of historical facts proportional to sequence length (Recall). We formalize this trade-off within an Online Sequence Processor abstraction that unifies Transformers, state space models, linear recurrent networks, and their hybrids. Using the Data Processing Inequality and Fano's Inequality, we prove that any model satisfying Efficiency and Compactness can recall at most O(poly(d)/log V) key-value pairs from a sequence of arbitrary length, where d is the model dimension and V is the vocabulary size. We classify 52 architectures published before March 2026 into the triangle, showing that each achieves at most two of the three properties and that hybrid architectures trace continuous trajectories in the interior. Experiments on synthetic associative recall tasks with five representative architectures validate the theoretical bound: empirical recall capacity lies strictly below the information-theoretic limit, and no architecture escapes the triangle.