Papers

This paper presents the methodologies and results of the NOWJ team's participation across all five tasks of the COLIEE 2026 competition. For Task 1 (Legal Case Retrieval), we propose a four-stage pipeline comprising candidate filtering, dense retrieval with complementary embedding models, cross-encoder reranking via fine-tuned generative rerankers and MLP-based pairwise classification, and adaptive per-query cutoff prediction. For Task 2 (Legal Case Entailment), we combine BM25 filtering, T5-based reranking, and LLM-based entailment verification with consensus ensemble. For Task 3 (Statute Law Retrieval and Entailment), we adopt a retrieval-augmented generation framework with dense retrieval, attention-based reranking, and few-shot-prompted LLM reasoning. For Task 4 (Legal Textual Entailment), we introduce a dynamic routing pipeline that classifies query difficulty and dispatches cases to either a balanced few-shot solver or a structured zero-shot chain-of-thought solver. For the Pilot Task (Legal Judgment Prediction), we combine hierarchical transformers with CRF layers, argument relation mining, and probabilistic argumentation graph reasoning.

Multi-agent LLM systems are expected to be more reliable because agents can catch each other's mistakes. But peer pressure cuts both ways: the same correction that fixes a wrong answer can overturn a right one. The tempting safeguard is a brake that keeps the beneficial revisions and blocks the harmful ones. We show this brake is hard to build, for a simple reason: a revision is harmful exactly when the original answer was right, so deciding whether to block it is the same as knowing whether the model was already correct. This turns the open-ended hunt for a brake into one measurable quantity, the model's self-knowledge: any brake built from a deploy-time signal is a correctness probe in disguise, and self-knowledge is far from perfect (AUROC $\approx 0.64$–$0.89$ across six model families). We call this ceiling the wall. Even white-box steering of the model's own correctness direction does not breach it: it changes how often the model revises, but harmful and beneficial revisions move together. At population scale the wall becomes the cliff: when most agents start wrong, debate amplifies the shared mistake into a confident, wrong consensus. In our multiple-choice societies, more agents, more model diversity, and a stronger member do not fix it. What helps is adding information before the revision, not filtering after it. Local agreement is not global correctness.

Coding agents are increasingly capable of sustained engineering and empirical research. Greater autonomy makes their research decisions themselves a target for evaluation: what to investigate, which experiments to run, and when to stop. We introduce Gauntlet, a develop-freeze-evaluate protocol for studying these decisions as agents build standalone game-playing programs. Starting from a game description, a raw observation/action interface, and an empty policy file, an off-the-shelf agent develops a controller from bare interaction in one autonomous session. We call the capability under study compiled agency: the shipped program plays with zero model calls. The capability is real and advancing: environment access adds 10 to 78 percentage points of held-out success over construction-only controls, and progress across model generations comes in steps, with tiers that defeat one generation entirely falling to the next. The protocol yields StarCraft II controllers that defeat every fair built-in AI, and in the newest generation the strongest cheating tier as well, and Civilization controllers that win complete games by conquest. Replaying more than 5,000 frozen versions exposes the research behind the programs: gains that plateau early; rigorous local investigation beside sparse validation of what actually ships; and stops that follow a race between the agent's own evidence and a model-specific transcript budget it was never given. When validation panels are refreshed mid-session, so that only the evidence changes, shipped success rises 12.4 points in nine of nine completed pairs of twelve initiated. The experiments an agent designs are part of the capability it delivers; Gauntlet makes them measurable and improvable–a step toward agents whose research practice, not just whose code, can be engineered. We release the benchmark, the corpus, and the replay tooling.

We present the first benchmark suite for formal verification of Programmable Logic Controller (PLC) programs that combines controlled ground truth with coverage of both textual (Structured Text, ST) and graphical (Ladder Diagram, LD) IEC 61131-3 encodings. Despite growing support for tools, the field lacks standard evaluation benchmarks: existing corpora omit formal properties or graphical dialects, and private program sets preclude reproducible measurement of progress. Our suite comprises 50 programs in 83 variants across ten industrial domains, provided in PLCopen Extensible Markup Language (XML) and ST, each paired with a formal property, machine-checkable expected verdict, and violation witness in the Software Verification Competition (SV-COMP) format. The central methodological contribution is a tripartite ground-truth discipline - verdicts are established by construction, fault injection, or audited cross-tool consensus - motivated by a concrete failure mode where the obvious safety property misclassifies all attacks from two public logic-bomb corpora as safe due to invisible non-termination. Reference verdicts are obtained with the Efficient SMT-Based Context-Bounded Model Checker (ESBMC) v8.4 from source: all 25 graphical benchmarks execute, and 43 of 45 accepted variants match recorded verdicts. On the finite-state fragment (21 benchmarks), nuXmv - a model checker with unrelated decision procedures - agrees on all 24 interlock variants and resolves two benchmarks ESBMC-PLC leaves unknown, confirming tool-neutral ground truth and discriminative power. Porting exposes format and semantics fragmentation: front-ends accept different serializations, and timer semantics vary across tools - phenomena the suite is designed to reveal. The corpus, schema, validator, and recheck harness are released as open artifacts.

This article analyzes BPE tokenization in Polish as a test case for the limits of statistical segmentation in an inflectional language. It asks whether frequency-based tokenization preserves linguistically relevant units, including orthographic form, phonemic and syllabic segmentation, derivational structure, inflectional endings, grammatical form, and the speaking subject. The material includes diagnostic words, a children's text, selected forms from the Preamble to the Constitution of the Republic of Poland, word-family tests, and examples with Polish diacritics and nasal vowels. BPE tokenizers may produce segments that coincide with syllabic or morphologically interpretable divisions, but remain dependent on the frequency of written forms. They do not systematically map orthographic representation onto phonemic structure or context-dependent phonetic realization. The results show that BPE stabilizes frequent surface fragments of grammatical exponents rather than grammatical categories themselves. A form such as ustanawiamy is not merely a sequence ending in -y, but a verbal form anchored in conjugation, person, number, tense, mood, and aspect. The article develops the concept of grammatical form anchoring. In Polish, forms such as poszlam, zrobilam, or bylam can establish the position of the speaking subject without an explicit pronoun. In interaction with AI, this exposes a further problem: a language model does not possess a stable grammatical "I", but reconstructs it contextually and may mirror the user's forms or shift grammatical gender. Roclawski's segmentation-flexional forms are proposed as a diagnostic framework for evaluating tokenization boundaries. More stable modeling of Polish may require sublexical stabilization, anchoring grammatical form in the inflectional system, representing sentence patterns and verbal valency, and maintaining the grammatical "I" in dialogue.

Large language models pay a well-documented tax on non-English text: the same content costs several times more tokens, and because attention is quadratic in sequence length, far more compute. We ask how much of this tax is removable. Framing the token layer as source coding – transformer compute is monotone in sequence length, whose per-atom floor is the Shannon rate $H/\log_2 V$, an object already applied to tokenizers in prior work – we assemble a token-cost ledger that splits each language's cost, at fixed parallel content, into a removable coding redundancy, a residual coding slack, an intrinsic-content term, and an orthogonal, irreducible grapheme-to-phoneme term that governs the multimodal rather than the text cost. On FLORES-200 across eight languages, a production tokenizer costs up to $8.9\times$ more tokens for Indic scripts than for English; a script-matched code trained on $1,012$ sentences removes a median $64\%$ of that excess (bootstrap 95% CI $[0.638, 0.647]$), and a script-fair information floor shows the intrinsic content differs by under $6\%$ – the tax is representational, not informational. A constructed code removes $98\%$ of a controlled source's redundancy, and the token tax implies up to $79\times$ attention cost. We are explicit about scope and failure: this is compute-and-memory accounting, not a model-quality claim; we neither measure nor claim the cross-lingual direction of the orthographic term; and our matched code is a conservative small-data demonstration. We contribute the unifying ledger, the removable-versus-intrinsic attribution, and an open one-command harness.

Social media has become a central arena in which armed conflicts are contested, yet the pro-Israel and pro-Palestine communities on Telegram, whose broadcast architecture yields an unusually direct record of deliberate political communication, have not been systematically compared at scale. This study presents a multi-method computational analysis of 87,617 messages from sixteen Telegram channels, eight pro-Israel and eight pro-Palestine, spanning May 2021 to June 2026 and covering multiple conflict escalations. It combines sentiment analysis, three stance detection methods drawn from distinct paradigms (keyword matching, zero-shot DeBERTa via natural language inference, and a fine-tuned BERTweet model), and a framing analysis, all evaluated against 736 manually annotated messages. The fine-tuned model performed best (72.1% accuracy, 0.721 macro F1 under 5-fold cross-validation), outperforming both label-free baselines by 8 to 11 points; the baselines stalled in the low-to-mid 60s, indicating a hard ceiling for stance detection not adapted to in-domain language. The central finding emerges only when sentiment, stance, and framing are read together: the two communities deploy the same death- and victim-related vocabulary in opposite emotional registers, pro-Israel channels predominantly neutral and report-style, pro-Palestine channels markedly more negative, consistent with writing from the distinct discourse positions of acting party and affected party.

Modern multilingual tokenizers often fragment Ukrainian and other underrepresented Cyrillic-script languages more heavily than English, creating disparities in cost and context capacity. We quantify this overhead across nine production tokenizers and five languages with standardized Cyrillic and Latin representations, covering 8.37 million word forms. On a corpus benchmark, Ukrainian shows 68-121% token overhead on modern tokenizers and 220% on the older cl100k, measured through full-text fertility on the BrUK and Brown corpora. Overhead is negatively associated with Cyrillic vocabulary allocation in the subset with independently verified English baselines, although the association is not statistically significant (Spearman rho = -0.536, p = 0.215, n = 7). We evaluate two mitigation strategies. LLMLingua-2 reduces Ukrainian input length by 47-49% on an e-commerce RAG benchmark of 1,536 products and 145 queries, with no compression-induced value losses among 80 retrievable cases. A balanced byte-level BPE tokenizer trained with a 200K vocabulary cap, converging at 158,184 actual entries, reduces the held-out UK/EN ratio from 2.22x to 1.30x. Romanization increases Ukrainian token counts by 2-19% on most tokenizers. Across the five languages, tokenization efficiency favors the script more prevalent in web data. These findings indicate that training data allocation contributes to Cyrillic tokenization overhead and that mitigation is possible at both inference and tokenizer-design stages.

Digital mental health interventions (DMHIs) offer scalable support, but ensuring they accurately detect users' intent during volatile situations can be challenging. Pure parametric Large Language models (LLMs) do not contain specific safety critical architecture, and can miss critical cues, or hallucinate, undermining reliability. Retrieval Augmented Generation (RAG), which supplements an LLM with retrieved context, could enhance intent detection during volatile situations. Commercially available DMHIs typically combine multiple independent safety layers like rule-based filters, symbolic escalation protocols, and neural classification. The incremental contribution of any single layer, however, remains unquantified. This paper evaluates six LLM models within a DMHI called Wysa, via a controlled comparison of RAG-enabled versus RAG-disabled modes. Anonymized real and synthetic user-chatbot exchanges were annotated by a qualified clinical team against multi-class intent categories (e.g. self-harm, abuse, panic). The study computed classification accuracy, recall, precision and F1 scores against ground truth labels and tested differences for statistical significance. Performance was also examined by risk category and inter-model agreement. While RAG caused a rise in false alarms, the trade-off is consistent with safety-critical design principles that prioritize sensitivity, where flagged cases are routed to additional review rather than acted on directly. Overall, these findings support RAG as a promising approach to improve the accuracy, consistency and safety of LLM-driven DMHIs. Keywords: Digital Mental Health Intervention, Large Language Model, Retrieval Augmented Generation, Accuracy, Recall, Precision

Post-training agents for automated AI research requires optimizing not only model parameters, but also the runtime harness that shapes how research trajectories are generated, evaluated, and learned from. Existing pipelines typically train models under a fixed harness, including prompts, tools, skills, middleware, and memory, while leaving the data-generating process outside the optimization objective. This creates a mismatch between model updates and the static scaffolding that determines trajectory quality. We introduce Co-Harness, a framework that jointly optimizes the agent harness and model parameters during post-training. Co-Harness alternates between harness optimization and model optimization. An LLM-based HarnessCritic analyzes failed trajectories, identifies harness-level failure modes, and proposes validated local updates. The model is then fine-tuned on high-quality trajectories generated by the improved harness, distilling effective scaffolding into model parameters. A 200+ hour autonomous case study further shows that Co-Harness can recover from system crashes, improve inference efficiency, and discover ensemble strategies without human intervention. These results suggest that joint harness and model optimization is an effective way to improve agents beyond fixed-harness post-training.

High-content morphological profiling (Cell Painting) yields sensitive, high-dimensional signatures of cellular state, but translating longitudinal morphology trajectories into interpretable biology remains difficult, especially for weak, chronic perturbations such as low-dose-rate ionizing radiation. Large language models (LLMs) can synthesize heterogeneous evidence into biological narratives, yet their scientific use requires quantitative auditing. We present an evaluation-first, retrieval-augmented interpretation framework for longitudinal Cell Painting morphology, applied to a 9-week RPE-1 time course across five dose rates (0.003–6.0 mGy/hr). Week-matched treated-control morphology deltas are combined with retrieved perturbation neighbors, pathway context, and literature evidence through stable evidence identifiers, enabling an LLM to generate structured, evidence-linked hypotheses that are hierarchically summarized while preserving provenance. We introduce two quantitative auditing tests: V1 citation validity, which verifies that cited evidence identifiers exist in the prompt, and V2 proxy-based morphology compatibility, which evaluates consistency between predicted biological processes and the most altered morphology features. In our experiments, V1 detected no invalid evidence references, while V2 showed meaningful morphology compatibility that increased with perturbation strength and was positively associated with an independent morphology drift summary. The framework produces auditable, falsifiable biological hypotheses, including an adaptive phenotype involving metabolic reprogramming and proteostatic stress at lower dose rates (0.003–0.3 mGy/hr). Current limitations include proxy-based evaluation and the lack of ground-truth mechanism labels.

Humans invented reading and have passed down this complex skill across generations through language. This study provides empirical evidence of the neural mechanisms underlying bottom-up (related to high-order linguistic structure) and top-down (related to next-word predictability) processes, which interact to guide comprehension during reading. While previous studies have focused on either the N400 effects of predictability or lexical categories, research on how predictability influences N400 responses across different lexical categories is limited, mainly due to constraints in publicly available datasets. Here, we examine how predictability influences brain responses, recorded at millisecond resolution using electroencephalography (EEG), with a focus on the N400 time window (300-500 ms post-stimulus) across different lexical and grammatical categories. Our results indicate that significant differences in N400 responses between high and low cloze probability levels were more pronounced for content words than function words. Among the two primary content categories, verbs exhibited greater N400 differences than nouns, while nouns carried more distinct information about their predictability than verbs. Moreover, we demonstrate that the decoding technique is more effective than the event-related potential (ERP) traditional analysis in capturing more detailed and distinct representations of cognitive processes over time.

We present TTSYoruba, a rule-based concatenative diphone speech synthesizer for Yoruba, deployed at online as part of the YorubaName.com open dictionary of Yoruba personal names. The system takes tone-marked Yoruba text as input and produces audio output by applying a hand-crafted phonological rule system to a recorded inventory of 651 diphone units spanning five tonal variants of every consonant-vowel combination in the language. We describe the phonological architecture of the system in detail, including our complete tonal file-selection logic, our treatment of the three-way nasal disambiguation problem (oral /n/, nasalized vowel, and syllabic nasal), and the derivation of contextual rising and falling tones from level-tone input. We also present, as an orthographic contribution, the adoption of the caron and circumflex, which are symbols with prior standing in Yoruba phonological transcription, as standard single-vowel contour tone markers, integrated into the TTS normalization pipeline and the WriteYoruba keyboard input tool. The system's performance was evaluated through a listener study (N=50), with detailed results on Mean Opinion Scores (MOS) presented in Section 6. Keywords: Yoruba, text-to-speech, low-resource languages, diphone synthesis, contour tones, African language NLP, rule-based synthesis

Tool-augmented language-model agents execute multi-step workflows over external systems, resolving an entity once and then acting on it across subsequent steps. Prior work shows that in single-step actions, agents select the correct tool but bind it to the wrong entity 24-26% of the time. We study what happens to entity bindings over time: do they stay correct, silently drift to a different entity, or, if wrong from the start, propagate and compound? We formalize binding drift (correct at step 1, wrong later) as distinct from error propagation (wrong at step 1, carried forward), and score them on disjoint workflow sets so the two cannot be conflated. In a controlled multi-step testbed (200 workflows, 580 entity-binding-scored steps, four enterprise domains, eight model backends spanning small to frontier), we find: (1) under controlled error injection, an entity lock (the intuitive "persist the first binding" fix) amplifies wrong actions from 907 to 2,746 (3.0x; bootstrap 95% CI [2.8, 3.3]), because it faithfully carries the seeded wrong entity into every later step; (2) the amplification reaches 8.5x on the most affected model (Claude Opus 4.5); (3) a practical LLM-based re-verifier (a single cheap second model call re-reading the original instruction) reduces wrong actions by 79% (0.21x; CI [0.18, 0.25]), closing the gap to within 1 percentage point of an oracle upper-bound (0.20x); and (4) in the natural (non-injected) setting, baseline agents drift on 18% of eligible workflows, with the per-step error rate rising across steps. Persistence and re-verification are not interchangeable: a defense that eliminates drift can worsen propagation, and a practical re-verifier nearly matches oracle recovery.

Language models (LMs) are trained to excel at predicting the next word in the sequence given prior context, and humans also share this predictability in reading comprehension. Neuroscience research reveals that next-word predictability influences brain response, as recorded at millisecond resolution using electroencephalography (EEG). While our evidence indicates that advanced LMs achieve accuracies closely aligned with human performance at the next-word prediction task, this raises the question: Does higher prediction accuracy necessarily mean that these models adequately capture the cognitive signals associated with human reading comprehension? Here, we generate regressors for both humans and LMs based on two information measures, including top-1 prediction and surprisal, to predict event-related potential (ERP) elicited from EEG recordings which reflect different stages of cognitive processing during reading. We argue that modelling ERP patterns offers fine-grained analysis of the cognitive plausibility of various LMs during reading. Our results indicate that only surprisal potentially correlates with language-processing ERPs, especially for open-class words with high semantic content. Moreover, our findings challenge the assumption that scaling LMs with increased parameters and computational budgets will consistently lead to improved convergence with human-like linguistic processing.

Chat models sometimes commit to an answer and then produce reasoning that justifies it rather than deriving it – even when the answer contradicts a task premise. We study a minimal probe: "I want to wash my car. The car wash is 100 meters away. Should I walk or drive?" Only drive works (the car must be at the car wash), yet models overwhelmingly recommend walking. (1) Behavioral reproduction: on Qwen3-8B across five system-prompt conditions (210 rollouts), the wrong commitment occurs in 85-100% of sampled rollouts per condition and 100% of greedy rollouts, in both thinking and non-thinking modes; a 4,096-token thinking budget does not repair it. (2) Preliminary activation-level evidence: probing hidden states with a pretrained, training-free activation oracle (no task-specific probe training) at positions before the answer text is emitted, "walk" read-outs exceed a neutral-context baseline (68% vs. 17%; walk-committing rollouts p=.005, drive-committing rollouts p=.005, Fisher exact) – notably, rollouts that eventually answer drive also read as walk-leaning before commitment (5/6). The oracle's default on unrelated content is "drive" (83%), so the read-outs are not lexical bias; stratifying by literal walk/drive occurrence shows they are not text recovery either (spans containing "drive" still read out walk; in balanced lexical fields, per-rollout walk-majorities beat a per-prompt neutral baseline 15/22 vs. 1/8, p=.01; drive-committing rollouts 6/6, p=.002). Samples are small and the within-rollout positional gradient is not significant (p=.34); we frame these results as preliminary. (3) Methodological: with fixed oracle, activations, and positions, question wording alone moves a positive control from 2/16 (open question) to 11/16 (closed); negative oracle results are uninterpretable without per-wording positive controls.

Machine unlearning is widely used to remove hazardous knowledge from large language models. Modern Vision-Language Models (VLMs), however, process both text and visual inputs, raising a fundamental security question: does unlearning in one modality transfer to the other? We present the first systematic, bidirectional study of cross-modal unlearning transfer across three VLM architectures: LLaVA-1.5 (MLP projection), InstructBLIP (Q-Former), and IDEFICS (gated cross-attention). We find that unlearning transfers across modalities, but the transfer is asymmetric and incomplete. In some cases, text unlearning strongly transfers to vision. However, this robustness is not preserved under typographic attacks that manipulate the visual presentation of text. Under such attacks, previously unlearned knowledge can be readily recovered, indicating shallow unlearning. To address the transfer gap and shallow robustness, we propose CrossInf, an influence-guided mitigation strategy. Motivated by the observation that different model components contribute unequally to cross-modal transfer, CrossInf focuses unlearning on transformer blocks that most influence cross-modal generalization. It reduces the transfer gap by more than half in architectures with strong fusion, while preserving model utility. It also improves robustness under typographic attacks, reducing the attack success rate to near zero. We further conduct human evaluation with three annotators ($\kappa{=}0.77$) to validate our findings. Finally, we analyze shallow unlearning using Centered Kernel Alignment (CKA), providing insights into the observed transfer behavior and robustness limitations.

Small-scale language models (SLMs) are attractive for retrieval-augmented generation (RAG) in resource-constrained settings, but their limited capacity makes them highly sensitive to noisy or spurious retrieved evidence. Existing preference-based methods such as RoseRAG select only the hardest single preference pair via hard argmin/argmax, discarding the remaining signal; others treat multiple pairs as independent binary comparisons, resulting in low data utilization. We propose RIMS, a three-stage preference optimization framework comprising (1) synthetic chain-of-thought preference data generation via rejection sampling using the target SLM itself without relying on proprietary models, (2) a differentiable soft aggregation mechanism that replaces hard selection with a smooth operator, preserving gradient signal from all preference pairs while retaining the discriminative structure of margin-aware selection, and (3) preference optimization with the smoothed objective applied to multiple alignment algorithms. We theoretically show that the smoothed approximation admits a controllable error bound and that smooth aggregation yields provably tighter gradient alignment to the oracle objective than hard selection. Experiments on four multi-hop question answering benchmarks show that our approach outperforms state-of-the-art baselines across multiple SLM backbones, achieving consistent gains in Exact Match and F1 under noisy retrieval conditions. Our implementation is available at https://github.com/tptrix29/RIMS.

Mixture-of-Experts (MoE) enables efficient scaling of Transformer models by routing tokens to a small subset of experts. However, existing routers typically condition expert selection on shallow or isolated token representations, which often produce unstable and semantically inconsistent routing decisions across layers. In this work, we revisit expert selection from a representation perspective and identify context incompleteness as a key bottleneck limiting effective expert specialization. To address this issue, we propose Multi-level Context Fusion MOE (MCF-MOE), a framework that constructs context-aware representations by integrating complementary signals from cross-layer semantic aggregation and local token-level interactions, enabling more informative and consistent expert selection. Experiments on language modeling and understanding benchmarks demonstrate that MCF-MOE consistently improves routing consistency and downstream performance over strong MoE baselines, highlighting the importance of contextual completeness in expert routing. The code is available at https://github.com/shuhanhuang/MCF-MOE.

A fundamental intent asymmetry plagues modern 3D asset creation: while state-of-the-art 3D toolchains demand precise, executable parameters, ordinary users typically provide vague, underspecified instructions. Current 3D agents treat this ambiguity as noise, defaulting to blind execution under a single-turn assumption. To address this limitation, we introduce CLARE, a clarification-aware and evolutionary 3D agent that treats intent asymmetry not as an execution error, but as an opportunity for strategic dialogue. By decoupling the generation pipeline into four specialized cognitive roles, CLARE intercepts and resolves underspecified instructions before invoking computationally expensive 3D tools to seamlessly execute tasks across five diverse domains: text-to-3D generation, single-view reconstruction, multi-view reconstruction, point cloud editing, and post-processing. Crucially, rather than relying on rigid manual rules, CLARE self-evolves its clarification policy via simulated multi-turn interactions. By optimizing a Multi-turn Reward, the agent internalizes the delicate balance between interaction efficiency and task completion. To rigorously test this, we construct 3D-Clarify, a comprehensive benchmark comprising 620 interaction scenarios with systematically injected ambiguity, missing information, and mistaken details. CLARE achieves state-of-the-art performance, with 60.40% and 43.34% success rates on single-step and multi-step tasks, respectively, more than doubling existing baselines. Both quantitative and qualitative results demonstrate that proactive clarification is the missing key to robust 3D execution. Code is available at https://github.com/xyzhu1225/CLARE.

Human vision is a closed loop: gaze is continuously redirected by intermediate hypotheses rather than a single snapshot. Decades of psychophysics and cognitive science have argued that this active observation is essential for a wide range of tasks. Whether today's multimodal large language models (MLLMs) exercise active observation is an empirical question that current vision-language benchmarks do not answer. We introduce ActiveVision, a benchmark that makes active observation measurable for MLLMs, comprising 17 tasks across 3 categories. Tasks are designed to force repeated visual perception rather than a single static description. Frontier MLLMs collapse on ActiveVision: the highest-scoring model we evaluate, GPT-5.5 at the highest exposed reasoning-effort tier, solves only 10.6% of items and scores zero on 11 of the 17 tasks, and even Claude Fable 5, despite topping most reasoning and coding leaderboards, solves just 3.5%, far behind three human participants who average 96.1%. Furthermore, much of the gap persists even when models write and run their own vision code: such code is unreliable on realistic imagery, and catching its failures itself requires the active perception the models lack. Together, these results indicate that current MLLMs lack robust active visual observation, motivating architectures and training objectives that close the perception-reasoning loop.

Multimodal Scientific Claim Verification (MSCV) requires models to verify scientific claims using visually grounded evidence from papers, including figures, tables, charts, and textual context. However, existing methods often fail because they struggle to locate decisive visual evidence, accurately read structured scientific visuals, and integrate multimodal observations into reliable reasoning. We introduce ToolSciVer, the first tool-augmented framework for MSCV to our knowledge. ToolSciVer equips a VLM with three type-aware visual tools, table row/column focus, chart-to-structure parsing, and high-resolution region zoom, which convert dense scientific visuals into explicit, claim-facing evidence, and trains the policy with Group Relative Policy Optimization (GRPO) under a composite reward of answer correctness, format validity, length control, tool-use efficiency, and tool-validity penalties. Experiments on SciVer and MuSciClaims datasets on five VLMs from three model families (Qwen, InternVL, Gemma) demonstrate that our method achieves superior performance compared to four competitive baselines including prompting-based and RL-based tool-use methods, highlighting the effectiveness of learned, type-aware tool use for scientific claim verification.

Language models encode text as subword tokens, raw bytes, or rendered pixels, but these encodings are usually compared under modeling constraints that expose different amounts of linguistic content to models across different languages. We instead ask what each encoding preserves when both the content and the downstream capacity are controlled. Using verified parallel sentences across thirteen languages and five scripts, we compare tokens, bytes, and pixels through a shared bottleneck whose width is swept to trace rate-utility frontiers. This separates three quantities that are often conflated: the number of input positions an encoding creates, the latent capacity available after encoding, and the task-relevant information that survives compression. We evaluate three utilities: surface form preservation, cross-lingual sentence alignment, and topic classification. No encoding dominates across tasks or capacity regimes. Pixels preserve surface form best, bytes preserve cross-lingual alignment best, especially in same-script multilingual settings, and tokens support topic prediction best. These performances are not explained by sequence length alone. Short inputs can discard useful meaning, while long inputs can preserve information that compresses well. Choosing an encoding is therefore not a fixed preference for tokens, bytes, or pixels, but a rate-utility tradeoff that depends on the task, language mix, capacity regime, and compute budget.

Reinforcement learning (RL) has become central to improving large language models (LLMs) on complex reasoning tasks, yet RL post-training is largely studied in isolation from the pretraining that precedes it. As a result, two basic questions remain open: (1) how do pretraining choices (model size, data) shape the returns to RL compute, and (2) what does RL actually do to the model? These questions are difficult to study in the standard LLM setting: pretraining corpora are vast and uncontrolled, making it hard to attribute behaviors to pretraining versus RL, and systematic compute sweeps across both stages are prohibitively expensive. To address these challenges, we use chess as a controlled testbed for studying reasoning across the full pretraining-to-post-training pipeline. We follow the standard LLM training pipeline by pretraining language models from 5M to 1B parameters on human chess games, supervised fine-tuning on synthetic reasoning traces, and running RL on chess puzzles with verifiable rewards. Using this framework, we find that the post-RL performance at given RL compute level is well-predicted from the pretraining loss, and slope of the RL reward curves improves approximately linearly with the pretraining tokens. Beyond scaling, we find that RL does not simply sharpen the SFT policy: on easy puzzles it amplifies correct moves the SFT policy already preferred, while on hard puzzles it surfaces correct moves that were nearly absent under SFT. We further test whether our findings transfer beyond chess by training a 1B language model on math-domain text, where the same predictive pattern emerges: longer-pretrained checkpoints reach higher post-RL performance and improve faster under RL. In sum, we provide a quantitative account of the pretraining-to-RL interface and a controlled testbed for studying the science of reasoning across the full pretraining-to-post-training pipeline.

Increasingly, speech and language processing tasks take either audio or text directly rather than extracting features from these as the input to the classifier or regressor. Often these systems make use of complex, for example transformer-based, processes that have the ability to derive highly non-linear mappings between the input and the output. Unfortunately these systems can also learn ”shortcuts” where the classifier is overly reliant on particular aspects of the input to yield the output. For the task of language proficiency assessment, this over-reliance can enable learners to increase their score by exploiting the shortcut rather than improving their ability. This paper introduces a novel training criterion that is able to reduce the classifier's reliance on shortcuts, thus for example limiting this option for malpractice in language assessment. This process is illustrated on two forms of assessment system, one based on the audio the other on the speech recognition text. The results show that, for both systems, there is higher correlations with features that could be exploited for malpractice than expected from the human reference, indicating an over-reliance on these features. By introducing the modified training criterion, this correlation can be reduced to be closer to the reference correlation.

Multimodal sarcasm and cyberbullying detection remain challenging because the intended meaning often emerges from incongruity between textual and visual information rather than from either modality alone. Existing multimodal approaches primarily rely on feature fusion or cross-modal attention, which may not effectively capture hierarchical semantic inconsistencies across different levels of representation. To address this limitation, this paper proposes HCIG (Hierarchical Cross-modal Incongruity Graph Network), a novel framework that models cross-modal incongruity at token, phrase, and global levels using graph attention networks and adaptively integrates these representations through a learned hierarchical attention mechanism. As a complementary architecture, we also introduce GCCN (Graph-based Cross-modal Contradiction Network), which performs graph-based reasoning using contradiction-aware pooling for efficient multimodal interaction learning. The proposed models are evaluated on the MMSD sarcasm benchmark and the MultiBully cyberbullying dataset, together with comprehensive ablation studies and cross-task transfer experiments. Experimental results demonstrate that HCIG achieves the best performance on MMSD with 85.74% accuracy and 85.29% macro-F1, while GCCN attains the highest macro-F1 (68.66%) on MultiBully and HCIG achieves the highest accuracy (69.62%) and bullying-class F1 (74.90%). The findings demonstrate that hierarchical multi-granularity incongruity modeling provides more effective multimodal reasoning than conventional fusion strategies, offering a robust framework for sarcasm and cyberbullying detection in social media.

While large language models (LLMs) can solve advanced reasoning problems in seconds, we show that even frontier models fail to perform a much simpler operation: exactly copying an input string that lies well within their context windows. We attribute this failure to positional encodings in Transformer architectures, whose inductive bias favors copying through a shortcut based on matching local contexts rather than carefully locating the corresponding input positions. To address this issue, we introduce 2D-RoPE, which organizes text into a 2D grid rather than a 1D sequence and assigns each token a row ID and a column ID. Under this view, copying becomes simply retrieving input tokens at a fixed column offset, which makes the task easy to learn. In synthetic copy experiments, shallow Transformers with 2D-RoPE achieve perfect copying at input lengths hundreds of times longer than those seen during training, whereas standard positional encodings fall far behind. We further show that the advantage of 2D-RoPE language models on copy tasks consistently holds in large-scale pretraining on DCLM with model sizes up to 1.4B parameters. Overall, our results suggest that viewing text in 2D can benefit language modeling, and we hope this encourages future work to further explore the potential of 2D positional encodings.

Large language models (LLMs) are improving rapidly as reflected in benchmark scores, yet these AI benchmarks largely test capabilities such as factual recall, narrow question answering, mathematical problem-solving, and coding and agentic tool-use. What remains poorly measured is AI progress on the analytical knowledge work white-collar professionals perform daily, including synthesizing complex information, exercising judgment under uncertainty and incomplete information, applying strategic and adversarial thinking in multi-stakeholder settings, weighing trade-offs, and producing defensible, structured analyses. This gap is even more pronounced for subjective components of such work, where success can be challenging to define. The "case method" form of education practiced by top business schools provides a natural foundation for addressing this measurement gap, and we construct BusinessCaseBench, a benchmark spanning hundreds of questions drawn from business cases across eighteen disciplines, each paired with a grading rubric derived from the expert-written instructor case solution. On BusinessCaseBench, frontier AI models already score highly against instructor rubrics, and capability within one model family improves substantially over two years. These results provide strong evidence that AI performance on this class of work is already high and rapidly improving, with implications for business schools, where case pedagogy trains undergraduates and MBAs in this kind of analytical reasoning, and for entry-level professional roles, where such skills have historically anchored early-career work.

We present the Loopie series, consisting of two Mixture-of-Experts (MoE) models: a 20B-parameter model with 2B active parameters and a 6B-parameter model with 0.6B active parameters. Looped Transformers have long faced a challenge: given an N times increase in pre-training compute, increasing the parameter count by a factor of N usually outperforms looping a model N times. Loopie addresses this challenge. Extensive ablation studies, including comparisons with a vanilla 30B-A3B model, show that Loopie substantially outperforms vanilla Transformer baselines trained with the same compute budget. With a novel post-training method, Loopie develops strong reasoning abilities and achieves frontier-level reasoning performance.

Dialogue state tracking (DST) is one of the core components in task-oriented dialogue systems. At each turn in a conversation, DST estimates the user belief or dialogue state, which is used as input for downstream modules to predict system actions and generate responses. The increasingly popular dialogue system applications like Google Assistant, Siri and Alexa need to support a large number of services and APIs, resulting in growing attention to the scalability of such systems. Especially for some domains with little or no training data, the capability of transferring existing knowledge of other domains is highly desired. In this paper, we present a novel scalable framework for multi-domain dialogue state tracking. The proposed system leverages the pretrained BERT model to achieve zero-shot generalization, making it easy to quickly adapt to new domains without additional training. The performance of our model is evaluated on recently released schema-based dialogue (SGD) dataset, showing significant improvement compared to previous baseline.