Papers

Automated judgment document generation is a significant yet challenging legal AI task. As the conclusive written instrument issued by a court, a judgment document embodies complex legal reasoning. However, existing methods often oversimplify this complex process, particularly by omitting the “Pre-Judge” phase, a crucial step where human judges form a preliminary conclusion. This omission leads to two core challenges: 1) the ineffective acquisition of foundational judicial elements, and 2) the inadequate modeling of the Pre-Judge process, which collectively undermine the final document's legal soundness. To address these challenges, we propose \textit{Judicial Unified Synthesis Through Intermediate Conclusion Emulation} (JUSTICE), a novel framework that emulates the “Search $\rightarrow$ Pre-Judge $\rightarrow$ Write” cognitive workflow of human judges. Specifically, it introduces the Pre-Judge stage through three dedicated components: Referential Judicial Element Retriever (RJER), Intermediate Conclusion Emulator (ICE), and Judicial Unified Synthesizer (JUS). RJER first retrieves legal articles and a precedent case to establish a referential foundation. ICE then operationalizes the Pre-Judge phase by generating a verifiable intermediate conclusion. Finally, JUS synthesizes these inputs to craft the final judgment. Experiments on both an in-domain legal benchmark and an out-of-distribution dataset show that JUSTICE significantly outperforms strong baselines, with substantial gains in legal accuracy, including a 4.6% improvement in prison term prediction. Our findings underscore the importance of explicitly modeling the Pre-Judge process to enhance the legal coherence and accuracy of generated judgment documents.

Contextual information at inference time, such as demonstrations, retrieved knowledge, or interaction history, can substantially improve large language models (LLMs) without parameter updates, yet its theoretical role remains poorly understood beyond specific settings such as in-context learning (ICL). We present a unified theoretical framework for analyzing the effect of arbitrary contextual information in Transformer-based LLMs. Our analysis characterizes contextual influence through output error dynamics. In a single-layer Transformer, we prove that the context-conditioned error vector decomposes additively into the baseline error vector and a contextual correction vector. This yields necessary geometric conditions for error reduction: the contextual correction must align with the negative baseline error and satisfy a norm constraint. We further show that the contextual correction norm admits an explicit upper bound determined by context-query relevance and complementarity. These results extend to multi-context and multi-layer Transformers. Experiments across ICL, retrieval-augmented generation, and memory evolution validate our theory and motivate a principled context selection strategy that improves performance by $0.6\%$.

With the continuous progress of digitization in Chinese judicial institutions, a substantial amount of electronic legal document information has been accumulated. To unlock its potential value, entity and relation extraction for legal documents has emerged as a crucial task. However, existing methods often lack domain-specific knowledge and fail to account for the unique characteristics of the judicial domain. In this paper, we propose an entity and relation extraction algorithm based on hypergraph neural network (Legal-KAHRE) for drug-related judgment documents. Firstly, we design a candidate span generator based on neighbor-oriented packing strategy and biaffine mechanism, which identifies spans likely to contain entities. Secondly, we construct a legal dictionary with judicial domain knowledge and integrate it into text encoding representation using multi-head attention. Additionally, we incorporate domain-specific cases like joint crimes and combined punishment for multiple crimes into the hypergraph structure design. Finally, we employ a hypergraph neural network for higher-order inference via message passing. Experimental results on the CAIL2022 information extraction dataset demonstrate that our method significantly outperforms existing baseline models.

Whether Reinforcement Learning with Verifiable Rewards (RLVR) endows Large Language Models (LLMs) with new capabilities or merely elicits latent traces remains a central debate. In this work, we align with the former view, proposing a probabilistic framework where capability is defined by instance-level solvability. We hypothesize that the emergence of complex reasoning can be driven by sharpening atomic step probabilities, which enables models to overcome the exponential decay of success rates inherent in multi-step reasoning chains. Utilizing the Algebrarium framework, we train models exclusively on single-step operations and evaluate their performance on unseen multi-step tasks. Our empirical results confirm that: (1) RLVR incentivizes the exploration of previously inaccessible solution paths by amplifying the model's existing skills; (2) composite performance is strictly governed by the joint probability of atomic steps, evidenced by high Pearson correlation coefficients ($\rho \in [0.69, 0.96]$); and (3) RLVR, acting as a global optimizer, can cause specific skills to be sacrificed to maximize aggregate reward. Our work offers a novel explanation for emergent abilities in RLVR, suggesting that the iterative optimization of solvable problems enables models to develop the capabilities to tackle previously unsolvable scenarios.

Elucidating the language-brain relationship requires bridging the methodological gap between the abstract theoretical frameworks of linguistics and the empirical neural data of neuroscience. Serving as an interdisciplinary cornerstone, computational neuroscience formalizes the hierarchical and dynamic structures of language into testable neural models through modeling, simulation, and data analysis. This enables a computational dialogue between linguistic hypotheses and neural mechanisms. Recent advances in deep learning, particularly large language models (LLMs), have powerfully advanced this pursuit. Their high-dimensional representational spaces provide a novel scale for exploring the neural basis of linguistic processing, while the "model-brain alignment" framework offers a methodology to evaluate the biological plausibility of language-related theories.

Language enables humans to share knowledge, reason about the world, and pass on strategies for survival and innovation across generations. At the heart of this process is not just the ability to communicate but also the remarkable flexibility in how we can express ourselves. We can express the same thoughts in virtually infinite ways using different words and structures - this ability to rephrase and reformulate expressions is known as paraphrase. Modeling paraphrases is a keystone to meaning in computational language models; being able to construct different variations of texts that convey the same meaning or not shows strong abilities of semantic understanding. If computational language models are to represent meaning, they must understand and control the different aspects that construct the same meaning as opposed to different meanings at a fine granularity. Yet most existing approaches reduce paraphrasing to a binary decision between two texts or to producing a single rewrite of a source, obscuring which linguistic factors are responsible for meaning preservation. In this thesis, I propose that decomposing paraphrases into their constituent linguistic aspects (paraphrase types) offers a more fine-grained and cognitively grounded view of semantic equivalence. I show that even advanced machine learning models struggle with this task. Yet, when explicitly trained on paraphrase types, models achieve stronger performance on related paraphrase tasks and downstream applications. For example, in plagiarism detection, language models trained on paraphrase types surpass human baselines: 89.6% accuracy compared to 78.4% for plagiarism cases from Wikipedia, and 66.5% compared to 55.7% for plagiarism of scientific papers from arXiv. In identifying duplicate questions on Quora, models trained with paraphrase types improve over models trained on binary pairs. Furthermore, I demonstrate that...

In light of increasing polarization and political violence, understanding the psychological roots of extremism is increasingly important. Prior research shows that identity fusion predicts willingness to engage in extreme acts. We evaluate the Cognitive Linguistic Identity Fusion Score, a method that uses cognitive linguistic patterns, LLMs, and implicit metaphor to measure fusion from language. Across datasets from the United Kingdom and Singapore, this approach outperforms existing methods in predicting validated fusion scores. Applied to extremist manifestos, two distinct high-fusion pathways to violence emerge: ideologues tend to frame themselves in terms of group, forming kinship bonds; whereas grievance-driven individuals frame the group in terms of their personal identity. These results refine theories of identity fusion and provide a scalable tool aiding fusion research and extremism detection.

There are two widely held characterizations of human semantic category systems: (1) they form convex partitions of conceptual spaces, and (2) they are efficient for communication. While prior work observed that convexity and efficiency co-occur in color naming, the analytical relation between them and why they co-occur have not been well understood. We address this gap by combining analytical and empirical analyses that build on the Information Bottleneck (IB) framework for semantic efficiency. First, we show that convexity and efficiency are distinct in the sense that neither entails the other: there are convex systems which are inefficient, and optimally-efficient systems that are non-convex. Crucially, however, the IB-optimal systems are mostly convex in the domain of color naming, explaining the main empirical basis for the convexity approach. Second, we show that efficiency is a stronger predictor for discriminating attested color naming systems from hypothetical variants, with convexity adding negligible improvement on top of that. Finally, we discuss a range of empirical phenomena that convexity cannot account for but efficiency can. Taken together, our work suggests that while convexity and efficiency can yield similar structural observations, they are fundamentally distinct, with efficiency providing a more comprehensive account of semantic typology.

Reinforcement Learning with Verifiable Rewards~(RLVR) has become a prominent paradigm to enhance the capabilities (i.e.\ long-context) of Large Language Models~(LLMs). However, it often relies on gold-standard answers or explicit evaluation rubrics provided by powerful teacher models or human experts, which are costly and time-consuming. In this work, we investigate unsupervised approaches to enhance the long-context capabilities of LLMs, eliminating the need for heavy human annotations or teacher models' supervision. Specifically, we first replace a few paragraphs with special placeholders in a long document. LLMs are trained through reinforcement learning to reconstruct the document by correctly identifying and sequencing missing paragraphs from a set of candidate options. This training paradigm enables the model to capture global narrative coherence, significantly boosting long-context performance. We validate the effectiveness of our method on two widely used benchmarks, RULER and LongBench~v2. While acquiring noticeable gains on RULER, it can also achieve a reasonable improvement on LongBench~v2 without any manually curated long-context QA data. Furthermore, we conduct extensive ablation studies to analyze the impact of reward design, data curation strategies, training schemes, and data scaling effects on model performance. We publicly release our code, data, and models.

Despite rapid progress in MLLMs, visual spatial reasoning remains unreliable when correct answers depend on how a scene would appear under unseen or alternative viewpoints. Recent work addresses this by augmenting reasoning with world models for visual imagination, but questions such as when imagination is actually necessary, how much of it is beneficial, and when it becomes harmful, remain poorly understood. In practice, indiscriminate imagination can increase computation and even degrade performance by introducing misleading evidence. In this work, we present an in-depth analysis of test-time visual imagination as a controllable resource for spatial reasoning. We first study when static visual evidence is sufficient, when imagination improves reasoning, and how excessive or unnecessary imagination affects accuracy and efficiency. To support this analysis, we then introduce AVIC, an adaptive test-time framework with world models that explicitly reasons about the sufficiency of current visual evidence before selectively invoking and scaling visual imagination. Finally, to further learn this gating and planning behavior without any annotation of when and how much to imagine, we introduce AVIC-R, which trains the policy via GRPO from QA-correctness rewards and penalties by imagination cost. Across spatial reasoning benchmarks (SAT, MMSI) and an embodied navigation benchmark (R2R), our results reveal clear scenarios where imagination is critical, marginal, or detrimental, and show that selective control can match or outperform fixed imagination strategies with substantially fewer world-model calls and language tokens. Our AVIC-R surpasses strong proprietary baselines including GPT-4o and GPT-4.1 while invoking the world model less often. Overall, our findings highlight the importance of analyzing and controlling test-time imagination for efficient and reliable spatial reasoning.

Although computer-use agents (CUAs) hold significant potential to automate increasingly complex OS workflows, they can demonstrate unsafe unintended behaviors that deviate from expected outcomes even under benign input contexts. However, exploration of this risk remains largely anecdotal, lacking concrete characterization and automated methods to proactively surface long-tail unintended behaviors under realistic CUA scenarios. To fill this gap, we introduce the first conceptual and methodological framework for unintended CUA behaviors, by defining their key characteristics, automatically eliciting them, and analyzing how they arise from benign inputs. We propose AutoElicit: an agentic framework that iteratively perturbs benign instructions using CUA execution feedback, and elicits severe harms while keeping perturbations realistic and benign. Using AutoElicit, we surface hundreds of harmful unintended behaviors from state-of-the-art CUAs such as Claude 4.5 Haiku, Claude 4.5 Opus, and Operator. We further evaluate the transferability of human-verified successful perturbations, identifying persistent susceptibility to unintended behaviors across various other frontier CUAs. This work establishes a foundation for systematically analyzing unintended behaviors in realistic computer-use settings.

Retrieval-Augmented Generation (RAG) often struggles with knowledge conflicts, where model-internal parametric knowledge overrides retrieved evidence, leading to unfaithful outputs. Existing approaches are often limited, relying either on superficial decoding adjustments or weight editing that necessitates ground-truth targets. Through layer-wise analysis, we attribute this failure to a parametric suppression phenomenon: specifically, in deep layers, certain FFN layers overwrite context-sensitive representations with memorized priors. To address this, we propose CoRect (Context-Aware Logit Contrast for Hidden State Rectification). By contrasting logits from contextualized and non-contextualized forward passes, CoRect identifies layers that exhibit high parametric bias without requiring ground-truth labels. It then rectifies the hidden states to preserve evidence-grounded information. Across question answering (QA) and summarization benchmarks, CoRect consistently improves faithfulness and reduces hallucinations compared to strong baselines.

Scaling large language models by increasing parameters and training data is increasingly constrained by limited high-quality corpora and rising communication costs. This work explores an alternative axis: increasing per-token computation without expanding parameters, by internalizing latent Chain-of-Thought (CoT) into pretraining. We propose Pretraining with Token-Level Adaptive Latent CoT (adaptive latent CoT), where the model generates a variable-length latent CoT trajectory before emitting each token – allocating longer trajectories to difficult tokens and shorter (or even zero) trajectories to easy ones. Importantly, this behavior emerges naturally from one-stage pretraining on general text and reduces computation in both training and inference via token-wise adaptive halting. Experiments with Llama architectures show that adaptive latent CoT consistently improves language modeling perplexity and broad downstream accuracy, even with fewer training FLOPs than prior recurrent baselines.

Molecule generation and optimization is a fundamental task in chemical domain. The rapid development of intelligent tools, especially large language models (LLMs) with powerful knowledge reserves and interactive capabilities, has provided new paradigms for it. Nevertheless, the intrinsic challenge for LLMs lies in the complex implicit relationship between molecular structure and pharmacological properties and the lack of corresponding labeled data. To bridge this gap, we propose DrugR, an LLM-based method that introduces explicit, step-by-step pharmacological reasoning into the optimization process. Our approach integrates domain-specific continual pretraining, supervised fine-tuning via reverse data engineering, and self-balanced multi-granular reinforcement learning. This framework enables DrugR to effectively improve key ADMET properties while preserving the original molecule's core efficacy. Experimental results demonstrate that DrugR achieves comprehensive enhancement across multiple properties without compromising structural similarity or target binding affinity. Importantly, its explicit reasoning process provides clear, interpretable rationales for each optimization step, yielding actionable design insights and advancing toward automated, knowledge-driven scientific discovery. Our code and model checkpoints are open-sourced to foster future research.

Humans align to one another in conversation – adopting shared conventions that ease communication. We test whether LLMs form the same kinds of conventions in a multimodal communication game. Both humans and LLMs display evidence of convention-formation (increasing the accuracy and consistency of their turns while decreasing their length) when communicating in same-type dyads (humans with humans, AI with AI). However, heterogenous human-AI pairs fail – suggesting differences in communicative tendencies. In Experiment 2, we ask whether LLMs can be induced to behave more like human conversants, by prompting them to produce superficially humanlike behavior. While the length of their messages matches that of human pairs, accuracy and lexical overlap in human-LLM pairs continues to lag behind that of both human-human and AI-AI pairs. These results suggest that conversational alignment requires more than just the ability to mimic previous interactions, but also shared interpretative biases toward the meanings that are conveyed.

Open-ended learning frames intelligence as emerging from continual interaction with an ever-expanding space of environments. While recent advances have utilized foundation models to programmatically generate diverse environments, these approaches often focus on discovering isolated behaviors rather than orchestrating sustained progression. In complex open-ended worlds, the large combinatorial space of possible challenges makes it difficult for agents to discover sequences of experiences that remain consistently learnable. To address this, we propose Dreaming in Code (DiCode), an unsupervised environment design (UED) framework in which foundation models (large language models) synthesize executable environment code to scaffold learning toward increasing competence. In DiCode, "dreaming" takes the form of materializing code-level variations of the world. We instantiate DiCode in Craftax, a challenging open-ended reinforcement learning benchmark characterized by rich mechanics and long-horizon progression. Empirically, DiCode enables agents to acquire long-horizon skills, achieving a 17% improvement in mean return over the strongest baseline and non-zero success on late-game combat tasks where prior methods fail. Our results suggest that code-level environment design provides a practical mechanism for curriculum control, enabling the construction of intermediate environments that bridge competence gaps in open-ended worlds.

Inference-time steering offers a promising way to control language models (LMs) without retraining. However, standard approaches typically rely on activation addition, which inevitably alters the hidden-state magnitudes raising concerns about representation collapse and degraded open-ended generation. In this work, we explore Spherical Steering, a training-free primitive that resolves this trade-off through activation rotation. Rather than shifting activations with a fixed vector, our method rotates them along a geodesic toward a target direction, preserving signal integrity while steering toward the target concept. To further enhance adaptivity, we incorporate a confidence gate that dynamically modulates steering strength based on input uncertainty. Extensive experiments across multiple-choice benchmarks demonstrate that Spherical Steering significantly outperforms addition-based baselines (notably by +10% on TruthfulQA, COPA, and Storycloze), while simultaneously maintaining the model's general open-ended generation quality. This work highlights the value of geometric consistency, suggesting that norm-preserving rotation is a robust and effective primitive for precise inference-time control. The code is available at: https://github.com/chili-lab/Spherical-Steering.

LLM unlearning is essential for mitigating safety, copyright, and privacy concerns in pre-trained large language models (LLMs). Compared to preference alignment, it offers a more explicit way by removing undesirable knowledge characterized by specific unlearning datasets. In previous works, gradient ascent (GA) and its variants have shown promise for implementing unlearning, yet their untargeted nature results in unintended degradation of general capabilities, incomplete removal of knowledge, and the generation of incoherent responses, among many others. We argue that these issues stem from the absence of explicit guidance on what and how models should unlearn. To fill this gap, we introduce a novel unlearning target, reasoning-based unlearning target, which satisfies both the specified unlearning scope and the specified post-unlearning response. Building on this, we propose targeted reasoning unlearning (TRU), which leverages reasoning-based unlearning target as guidance. We employ the target using a cross-entropy supervised loss combined with a GA-based loss, enabling the model to learn reasoning ability for precise knowledge removal while preserving unrelated abilities. We evaluate TRU against strong baselines across multiple benchmarks and LLM backbones, and find that it achieves more reliable unlearning while preserving general capabilities. Moreover, TRU exhibits superior robustness under diverse attack scenarios, stemming from the reasoning ability learned through reasoning-based targets. Overall, our study establishes reasoning-augmented unlearning as a practical paradigm for reliable and explainable LLM unlearning.

This paper documents the architecture, development practices, and preserved artifacts of the Logovista English–Japanese machine translation system, a large, explicitly rule-based MT system that was developed and sold commercially from the early 1990s through at least 2012. The system combined hand-authored grammatical rules, a large central dictionary encoding syntactic and semantic constraints, and chart-based parsing with weighted interpretation scoring to manage extensive structural ambiguity. The account emphasizes how the system was extended and maintained under real-world usage pressures, including regression control, ambiguity management, and the limits encountered as coverage expanded. Unlike many rule-based MT systems described primarily in research settings, Logovista was deployed for decades and evolved continuously in response to practical requirements. The paper is intended as a technical and historical record rather than an argument for reviving rule-based MT, and describes the software and linguistic resources that have been preserved for potential future study.

Modern large language model (LLM) inference engines optimize throughput and latency under fixed decoding rules, treating generation as a linear progression in token time. We propose a fundamentally different paradigm: entropic\-time inference, where decoding is governed by the flow of uncertainty rather than token index. We introduce a self\-organizing inference architecture that jointly couples scheduling, attention sparsification, and sampling temperature under a unified entropy control objective. Our method extends vLLM with entropy-aware scheduling, entropic pruning of paged attention blocks, and adaptive temperature control that stabilizes generation near a target entropy regime. This transforms inference into a resource\-intelligent thermodynamic process that allocates computation where uncertainty reduction is maximized. We present a concrete systems design, pseudocode, and integration plan, demonstrating how entropy can serve as a first\-class control signal for scalable LLM inference.

Cold start scenarios present fundamental obstacles to effective recommendation generation, particularly when dealing with users lacking interaction history or items with sparse metadata. This research proposes an innovative hybrid framework that leverages Large Language Models (LLMs) for content semantic analysis and knowledge graph development, integrated with cognitive profiling based on VARK (Visual, Auditory, Reading/Writing, Kinesthetic) learning preferences. The proposed system tackles multiple cold start dimensions: enriching inadequate item descriptions through LLM processing, generating user profiles from minimal data, and dynamically adjusting presentation formats based on cognitive assessment. The framework comprises six integrated components: semantic metadata enhancement, dynamic graph construction, VARK-based profiling, mental state estimation, graph-enhanced retrieval with LLM-powered ranking, and adaptive interface design with iterative learning. Experimental validation on MovieLens-1M dataset demonstrates the system's capacity for personalized recommendation generation despite limited initial information. This work establishes groundwork for cognitively-aware recommendation systems capable of overcoming cold start limitations through semantic comprehension and psychological modeling, offering personalized, explainable recommendations from initial user contact.

How does the conversational past of large language models (LLMs) influence their future performance? Recent work suggests that LLMs are affected by their conversational history in unexpected ways. For instance, hallucinations in prior interactions may influence subsequent model responses. In this work, we introduce History-Echoes, a framework that investigates how conversational history biases subsequent generations. The framework explores this bias from two perspectives: probabilistically, we model conversations as Markov chains to quantify state consistency; geometrically, we measure the consistency of consecutive hidden representations. Across three model families and six datasets spanning diverse phenomena, our analysis reveals a strong correlation between the two perspectives. By bridging these perspectives, we demonstrate that behavioral persistence manifests as a geometric trap, where gaps in the latent space confine the model's trajectory. Code available at https://github.com/technion-cs-nlp/OldHabitsDieHard.

Epistemic Network Analysis (ENA) is a method for investigating the relational structure of concepts in text by representing co-occurring concepts as networks. Traditional ENA, however, relies heavily on manual expert coding, which limits its scalability and real-world applicability to large text corpora. Topic modeling provides an automated approach to extracting concept-level representations from text and can serve as an alternative to manual coding. To tackle this limitation, the present study merges BERTopic with ENA and introduces TopicENA, a topic-based epistemic network analysis framework. TopicENA substitutes manual concept coding with automatically generated topics while maintaining ENA's capacity for modeling structural associations among concepts. To explain the impact of modeling choices on TopicENA outcomes, three analysis cases are presented. The first case assesses the effect of topic granularity, indicating that coarse-grained topics are preferable for large datasets, whereas fine-grained topics are more effective for smaller datasets. The second case examines topic inclusion thresholds and finds that threshold values should be adjusted according to topic quality indicators to balance network consistency and interpretability. The third case tests TopicENA's scalability by applying it to a substantially larger dataset than those used in previous ENA studies. Collectively, these cases illustrate that TopicENA facilitates practical and interpretable ENA analysis at scale and offers concrete guidance for configuring topic-based ENA pipelines in large-scale text analysis.

Recently presented Token-Oriented Object Notation (TOON) aims to replace JSON as a serialization format for passing structured data to LLMs with significantly reduced token usage. While showing solid accuracy in LLM comprehension, there is a lack of tests against JSON generation. Though never present in training data, TOON syntax is simple enough to suggest one-shot in-context learning could support accurate generation. The inevitable prompt overhead can be an acceptable trade-off for shorter completions. To test this, we conducted a benchmark creating several test cases with regard to structural complexity, a validation pipeline, and comparing plain JSON generation vs structured output (via constrained decoding) JSON generation vs TOON one-shot in-context learning generation. JSON structured output was included to establish a minimum token budget baseline and to set a starting point for future experiments testing TOON constrained decoding inference enforcement. Key findings: TOON shows promising accuracy/token consumption ratio for in-domain generation tasks, though this advantage is often reduced by the "prompt tax" of instructional overhead in shorter contexts. Plain JSON generation shows the best one-shot and final accuracy, even compared with constrained decoding structured output, where the only significant advantage is the lowest token usage as a trade-off for slightly decreased accuracy overall and significant degradation for some models. Notably, for simple structures, this "lowest token usage" of constrained decoding outperformed even TOON, hinting that TOON enforcing via frameworks such as xgrammar may not yield the desired results. Furthermore, the results suggest a scaling hypothesis: TOON's true efficiency potential likely follows a non-linear curve, shining only beyond a specific point where cumulative syntax savings amortize the initial prompt overhead.

Large language models (LLMs) are increasingly used to generate executable outputs, JSON objects, and API calls, where a single syntax error can make the output unusable. Constrained decoding enforces validity token-by-token via masking and renormalization, but it can distort generation when the model assigns low probability mass to valid continuations, pushing decoding toward locally valid yet semantically incorrect trajectories. We propose Draft-Conditioned Constrained Decoding (DCCD), a simple two-step, training-free inference procedure that decouples semantic planning from structural enforcement: an unconstrained draft is generated first, and constrained decoding is then applied, conditioned on this draft, to guarantee validity. We analyze DCCD through a KL-projection view, showing that draft conditioning increases feasible mass and reduces the cumulative "projection tax" induced by hard constraints, with an optional best-of-$K$ draft selection. Across structured reasoning benchmarks, DCCD improves strict structured accuracy by up to +24 percentage points over standard constrained decoding (e.g., 15.2% to 39.0% on GSM8K with a 1B model), and enables smaller model pairs to match or exceed much larger constrained baselines, yielding substantial gains in parameter efficiency.

This article introduces the concept of the algorithmic unconscious to designate the set of structural determinations that operate within large language models (LLMs) without being accessible either to the model's own reflexivity or to that of its users. In contrast to approaches that reduce AI bias solely to dataset composition or to the projection of human intentionality, we argue that a significant class of biases emerges directly from the technical mechanisms of the models themselves: tokenization, attention, statistical optimization, and alignment procedures. By framing bias as an infrastructural phenomenon, this approach resolves a central theoretical ambiguity surrounding responsibility, neutrality, and correction in contemporary LLMs. Based on a comparative analysis of tokenization across a corpus of parallel sentences, we show that Arabic languages (Modern Standard Arabic and Maghrebi dialects) undergo a systematic inflation in token count relative to English, with ratios ranging from 1.6x to nearly 4x depending on the infrastructure (OpenAI, Anthropic, SentencePiece/Mistral). This over-segmentation constitutes a measurable infrastructural bias that mechanically increases inference costs, constrains access to contextual space, and alters attentional weighting within model representations. We relate these empirical findings to three additional structural mechanisms: causal bias (correlation vs causation), the erasure of minoritized features through dimensional collapse, and normative biases induced by safety alignment. Finally, we propose a framework for a technical clinic of models, grounded in the audit of tokenization regimes, latent space topology, and alignment systems, as a necessary condition for the critical appropriation of AI infrastructures.

Chart understanding is a quintessential information fusion task, requiring the seamless integration of graphical and textual data to extract meaning. The advent of Multimodal Large Language Models (MLLMs) has revolutionized this domain, yet the landscape of MLLM-based chart analysis remains fragmented and lacks systematic organization. This survey provides a comprehensive roadmap of this nascent frontier by structuring the domain's core components. We begin by analyzing the fundamental challenges of fusing visual and linguistic information in charts. We then categorize downstream tasks and datasets, introducing a novel taxonomy of canonical and non-canonical benchmarks to highlight the field's expanding scope. Subsequently, we present a comprehensive evolution of methodologies, tracing the progression from classic deep learning techniques to state-of-the-art MLLM paradigms that leverage sophisticated fusion strategies. By critically examining the limitations of current models, particularly their perceptual and reasoning deficits, we identify promising future directions, including advanced alignment techniques and reinforcement learning for cognitive enhancement. This survey aims to equip researchers and practitioners with a structured understanding of how MLLMs are transforming chart information fusion and to catalyze progress toward more robust and reliable systems.

Local governance meeting records are official documents, in the form of minutes or transcripts, documenting how proposals, discussions, and procedural actions unfold during institutional meetings. While generally structured, these documents are often dense, bureaucratic, and highly heterogeneous across municipalities, exhibiting significant variation in language, terminology, structure, and overall organization. This heterogeneity makes them difficult for non-experts to interpret and challenging for intelligent automated systems to process, limiting public transparency and civic engagement. To address these challenges, computational methods can be employed to structure and interpret such complex documents. In particular, Natural Language Processing (NLP) offers well-established methods that can enhance the accessibility and interpretability of governmental records. In this focus article, we review foundational NLP tasks that support the structuring of local governance meeting documents. Specifically, we review three core tasks: document segmentation, domain-specific entity extraction and automatic text summarization, which are essential for navigating lengthy deliberations, identifying political actors and personal information, and generating concise representations of complex decision-making processes. In reviewing these tasks, we discuss methodological approaches, evaluation metrics, and publicly available resources, while highlighting domain-specific challenges such as data scarcity, privacy constraints, and source variability. By synthesizing existing work across these foundational tasks, this article provides a structured overview of how NLP can enhance the structuring and accessibility of local governance meeting records.

When a language model asserts that "the capital of Australia is Sydney," does it know this is wrong? Models assert misconceptions with the same fluency as facts, so the question cannot be answered from output uncertainty. Truth-related signals are known to exist in the residual stream, but not their geometry: how many dimensions carry the signal, how simple a detector can be, and whether it transfers. We characterize this geometry across 11 models (124M-14B) and test it causally with activation steering, concept erasure, and distributed alignment search. The structure is simple: two class centroids in a 2-8 dimensional subspace match a trained linear probe, and 25 labeled examples recover 90% of full-data AUC on GPT-2. Steering shifts hallucination rates by 9.1 points on six models, erasure drops detection to chance, and distributed alignment search, the only method that bounds rank, localizes at most five causal dimensions. The internal advantage is regime-specific: probes far outperform P(True) and semantic entropy on adversarial misconceptions but tie on standard QA. Single-dataset probes transfer near-randomly until joint multi-dataset training restores 0.73-0.91 AUC. That centroid distance matches probe performance indicates class separation is a mean shift, making detection geometric rather than learned.

Dialogues are a predominant mode of communication for humans, and it is immensely helpful to have automatically generated summaries of them (e.g., to revise key points discussed in a meeting, to review conversations between customer agents and product users). Prior works on dialogue summary evaluation largely ignore the complexities specific to this task: (i) shift in structure, from multiple speakers discussing information in a scattered fashion across several turns, to a summary's sentences, and (ii) shift in narration viewpoint, from speakers' first/second-person narration, standardized third-person narration in the summary. In this work, we introduce our framework DIALSUMMER to address the above. We propose DIAL-SUMMER's taxonomy of errors to comprehensively evaluate dialogue summaries at two hierarchical levels: DIALOGUE-LEVEL that focuses on the broader speakers/turns, and WITHIN-TURN-LEVEL that focuses on the information talked about inside a turn. We then present DIAL-SUMMER's dataset composed of dialogue summaries manually annotated with our taxonomy's fine-grained errors. We conduct empirical analyses of these annotated errors, and observe interesting trends (e.g., turns occurring in middle of the dialogue are the most frequently missed in the summary, extrinsic hallucinations largely occur at the end of the summary). We also conduct experiments on LLM-Judges' capability at detecting these errors, through which we demonstrate the challenging nature of our dataset, the robustness of our taxonomy, and the need for future work in this field to enhance LLMs' performance in the same. Code and inference dataset coming soon.