Papers

Object hallucination in multimodal large language models arises when language priors and corpus co-occurrence bias outweigh the visual evidence, with nothing tying an object mention to the image. Most remedies intervene at decoding time, yet under a unified protocol their benefit is confined to short captions; supervised fine-tuning (SFT) on a detail-rich corpus lengthens captions, but over forty percent still name absent objects. This paper proposes Dual-Stream Cross-Anchor Correction (DSCC). Unlike work that post-processes decoding, DSCC injects object-level visual anchors into the language model itself during fine-tuning: a perception stream aligns object-level hidden states at an intermediate layer to frozen text anchors by a bidirectional contrastive objective; a cognition stream lets deeper layers query those anchors by cross-attention at every generation step; and a two-stage curriculum gate couples them, making evidence retrieval a structural constraint on generation. Under one backbone and one scoring protocol, experiments span long-caption hallucination, object-existence discrimination and cross-domain generalisation, with vanilla SFT on the same corpus and schedule as a length- and density-matched control separating the data effect from the architectural gain. DSCC alone reaches the long-caption, low-hallucination region: captions roughly 1.9 times the baseline length at 88.19% precision per object mention, the highest under a density-independent criterion. Ablations expose a synergy: the perception stream alone degrades precision yet reverses sign when stacked on the cognition stream. No universal superiority is claimed: three out-of-domain benchmarks yield a predictable, falsifiable domain-conditionality, the synergy being bound to the anchors' semantic domain and breaking on charts and illusions.

While Large Language Model (LLM) agents increasingly rely on long-term memory for persistent interactions, the retrieval mechanisms governing this memory are rarely treated as evolvable components. This static approach limits performance on heterogeneous memory queries, which often demand diverse evidence construction strategies. To address this, we introduce ERSkill, a retrieval-centric framework for evolving, skill-guided memory access. ERSkill compiles interaction histories into a structured memory store and represents retrieval behaviors as executable skills composed of fundamental primitives. At inference time, a trained router dynamically matches each query to a suitable retrieval skill to construct tailored evidence for answer generation. ERSkill co-evolves the skill set and the router during training. It employs an experience trie to efficiently record explored retrieval paths, alongside a double-frontier mechanism that separates oracle-side capability expansion from router-validated deployment. Experiments across multiple agent memory benchmarks demonstrate that ERSkill substantially outperforms strong non-evolving and evolving baselines. Notably, it improves the overall average across F1, BLEU-1, and LLM-judge scores by 31.3% with Qwen3-Next-80B-A3B-Instruct and by 21.4% with GPT-5.4-nano.

Mechanistic interpretability of large language models lacks spatially resolved, falsifiable tools for testing whether internal components are specialized for distinct cognitive operations. We adapt subtraction analysis, the standard framework of human neuroimaging, from biological brains to perturbed transformers, and apply the same logic to both substrates in parallel. Building on the Brain-LLM Unified Model (BLUM), which showed that layer-perturbed LLaVA-1.6-Vicuna-13B error profiles match the lesion patterns of aphasic patients, we develop PRISM (Perturbation-based Regional Interpretability through Subtraction Mapping). PRISM maps the seven clinical Philadelphia Naming Test categories, subtracts error classes pairwise, and treats each perturbation seed as a subject in a group analysis with threshold-free cluster enhancement along the layer axis. We run a structurally matched analysis on 213 chronic post-stroke aphasia patients using correlation-difference lesion-symptom mapping, and replicate both sides on held-out splits. The designs match in subject dimension (seeds, patients), spatial dimension (layers, atlas-parcellated cortex) and thresholding, but the contrast operator differs: a within-subject error-proportion difference for the LLM, a between-subject correlation difference for the cortex. Both substrates recover a robust phonemic-favoring dissociation, a deep layer cluster and a frontal-perisylvian cortical cluster, both replicating; the semantic-favoring direction is a consistently signed but non-significant trend on both. PRISM thus gives a falsifiable, spatially resolved test of functional-specialization claims in transformer language models. A confirmatory ROI-level intervention (PRISM Stage 3) licensing the strongest causal-mechanism claim is left to subsequent work.

Watermarking LLM-generated text is an important task for tracing its provenance. Existing LLM watermarks preserve provenance under editing, but this same robustness allows an adversary to alter critical content while retaining attribution, a vulnerability known as piggyback spoofing. We introduce an innovative watermark that jointly provides provenance and tamper evidence. It co-embeds a robust signal and a fragile signal into each generated token. The signals share the same mechanism but use independent keys and different seeding windows over normalized text, making one resilient to edits and the other sensitive to reader-visible changes. Multiple rounds of unbiased tournament reweighting preserve the expected generation distribution, while a periodic round-allocation pattern controls the trade-off between the two signals. At detection, their scores form a two-dimensional space supporting three decisions: Intact, Tampered, and No-Watermark. Across two large language models and two prompt datasets, our method demonstrates the highest tamper-detection rate among the evaluated methods while maintaining competitive attribution robustness and perplexity. Ablation studies show that reliable three-state detection requires a well-defined notion of intactness, co-embedding of the two signals, and complementary sensitivity to edits.

As urban residents demand higher-quality instant delivery services, demand forecasting, estimated time of arrival prediction, and next-stop prediction have become critical tasks in urban delivery systems. However, the three tasks rely on historical evidence with substantially different structures and granularities. Meanwhile, their prediction targets differ in output form, making it difficult to integrate retrieved evidence through a single fusion mechanism without causing mismatch or negative transfer. These issues can increase prediction errors and subsequently lead to imbalanced resource allocation, where some regions experience unmet demand. In contrast, others suffer from underutilization, ultimately degrading overall efficiency and service quality. To address these challenges, we propose TAGR, a Task-Adaptive Geo-Semantic Retrieval framework across heterogeneous urban delivery tasks. TAGR constructs an LLM-based geo-semantic shared prior and task-specific retrieval banks to share regional knowledge while preserving heterogeneous task-specific evidence. The LLM-derived regional representations encode geographic context and semantic relations among regions, providing transferable knowledge for all three tasks without requiring online LLM inference. It further introduces a task-adaptive retrieval augmentation module to accommodate heterogeneous output structures, thereby mitigating information mismatch and negative transfer caused by a uniform retrieval-fusion strategy. Experiments on three city datasets demonstrate that TAGR consistently outperforms state-of-the-art baselines across multiple tasks.

News headlines frame public issues both by what they select and by how they word it, yet computational framing work typically collapses these operations into a single score. We introduce a two-dimensional framework that separates salience framing, measured through four wording devices (loaded vocabulary, blame attribution, threat framing, rhetorical question), from selection framing, measured through outlet-level story-form and high-charge distributions. We build a 10,000-headline French supervision set using three LLM annotators with majority-vote resolution and human arbitration, validate the labels against two annotator-independent blind human studies, and apply the strongest classifier to 902,111 deduplicated headlines from 25 French outlets (2022-2025). Three main findings emerge. First, salience and selection divergence are positively correlated yet leave nearly half of outlet-level variance unexplained, populating interpretively distinct off-diagonal cells in a four-cell outlet typology. Second, default classification thresholds systematically inflate corpus-level salience estimates; a precision-floor recalibration protocol corrects this distortion. Third, group-mention analysis reveals sharply unequal salience contexts: headlines mentioning Jews, the Far-right, and Muslims carry the highest detected salience rates, which broad event-context composition does not fully explain (residuals are descriptive, not same-event causal estimates; per-group lexicon precision is reported alongside). To our knowledge, this is the largest framing-focused French headline audit to date; we release the supervision set, lexicons, and analysis code.

Post-training quantization and pruning rely on a small calibration corpus. Whether specialized domains such as finance require domain-matched calibration data remains unsettled. We argue that the answer depends on the task-level damage caused by compression rather than on domain mismatch. If compression preserves the target capability, changing the calibration corpus has little effect. If compression causes large losses, task-formatted calibration can recover part of the loss. We test this hypothesis across two model families, six compression configurations, three token-matched calibration corpora, and ten financial classification and numerical question-answering tasks. The results support this hypothesis. Quantization largely preserves task performance, and calibration choice has little effect in this case. Pruning reduces numerical QA accuracy by over 40 points. In these damaged settings, another generic corpus does not help, while FinMix, a mixture of financial task examples, recovers a large part of the loss. The link between damage and recovery holds across model families and scales. These findings support a practical rule. Measure task-specific compression damage first, and construct specialized calibration data only when the damage is large.

We present the ABAI submission to COLIEE 2026 Task 1, case law retrieval, together with a controlled study of why it underperformed. The task suppresses the cited passages themselves, which removes much of the lexical overlap a retriever would rely on. Our pipeline answers this with four independently trained stages: multi-view BM25 over citation-context windows with reciprocal rank fusion, neural reranking, graph-based features from entity communities and a graph attention network, and a LightGBM meta-learner over 34 features. Our best run reached F1=0.177 on the official test set, against a cross-validated 0.311, and we attributed that gap to a recall ceiling, temporal distribution shift, and threshold miscalibration. We then tested all three. Under leakage-free protocols threshold transfer costs 0.007 F1, decision quality is flat across chronological quartiles, and the official test queries are not measurably farther from the training manifold than training queries are from each other, in two independent embedding spaces. Decomposing the misses instead splits them exactly evenly between candidates never retrieved and candidates retrieved but ranked below the cut. Measuring the remedies for each half, BM25 length-normalisation tuning, an event-triple view, and full-content dense fusion lift top-200 recall by three to seven points, and citation-graph features add 0.014 F1 over eight seeds once own-citation leakage is removed, while per-query cutoff rules, a zero-shot reranker swap, and a date filter do not help. We also document four evaluation artifacts, each of which reversed a result once the protocol was corrected.

Vision-language models (VLMs) are increasingly used to interpret charts and generate natural-language explanations for socially consequential data. However, they may produce different narratives for the same chart when only the referenced social group changes, reinforcing stereotypes and misleading decisions. Despite these risks, no benchmark exists for systematically evaluating bias in chart interpretation across social dimensions. We introduce ChartBias, the first benchmark for auditing bias in VLM-based chart interpretation. ChartBias contains 820 manually curated real-world charts spanning six attributes: race, income, age, religion, immigration status, and gender, yielding 4,319 valid chart, attribute instances and 8,638 paired generations where the chart is fixed and only the group term is swapped. Across 12 proprietary and open-source VLMs, totaling 155,484 model responses, we find three widespread failure modes: narrative shift (same chart, different narratives), group hallucination (assigning a chart to a group without evidence), and preference polarity (favourable trends often linked to one group). We further propose a multi-agent mitigation framework that serves as a strong baseline by separating chart-grounded evidence extraction from group-conditioned generation and using a counterfactual judge to verify that group-driven differences are supported by the chart. The framework substantially reduces narrative shift while preserving chart-grounded reasoning. Our findings show that evaluating chart understanding requires measuring not only accuracy, but also fairness and consistency across social groups. We release ChartBias at https://github.com/vis-nlp/ChartBiasBench.

Large Language Models (LLMs) are widely used as programming assistants, yet it remains unclear whether and how user's demographic information impacts the technical quality of generated code. We conduct a large-scale empirical study of persona-induced bias in LLM-based code generation, focusing a proprietary model (Gemini 2.5 Pro) and an open-weight model (GPT-OSS-120B). Using 18 demographic personas spanning nationality, gender, and experience level, we compare persona-induced prompts against a neutral baseline. Across 35,000+ generated programs, we analyze demographic marker leakage in reasoning and responses, as well as differences in functional correctness, maintainability, code style, and security. Our results show that demographic cues are frequently reflected in LLM reasoning and outputs. Demographic markers appear in up to 65% of responses and 70% of reasoning traces, despite being semantically irrelevant to the tasks. On LiveCodeBench, persona prompting were associated with lower correctness scores of the Gemini model by an average of 1.54 percentage points, with one persona exhibiting a decrease of 3.6% (odds ratio = 0.51). In contrast, the accuracy of the GPT-OSS model improved by 3.4 - 5.7% across all personas (odds ratios = 1.8 - 3.0). Maintainability and code style metrics show statistically significant but negligible effect sizes (all Cliff's {\delta} < 0.15), and security vulnerabilities exhibit no systematic persona-specific patterns. Overall, our results show that the presence of demographic information about users is associated with measurable variation in LLM reasoning and code quality even in purely technical tasks, and that these effects hold across models. Our work highlights an under-examined risk in LLM-assisted software development.

Large language models can process increasingly long prompts, yet their ability to locate and use decisive evidence may degrade as irrelevant or confusable context is added. We formulate this phenomenon, which we call context poisoning, as extreme-value interference in attention: the decisive-evidence score is upper-bounded, while the maximum score among effective distractors grows with their number. Under a softmax retrieval abstraction, we derive a finite-sample upper bound showing that maintaining a fixed accuracy target above base rate requires the evidence margin to scale as $\Omega(\sqrt{\log N})$, where N denotes the effective distractor count rather than necessarily the raw context length. The analysis connects long-context degradation to score aliasing, positional aliasing, and softmax dilution. Controlled experiments show that retrieval accuracy decreases as total context grows in the presence of embedded hard negatives, that the same-format condition produces the largest observed accuracy drop among the tested distractor constructions at fixed context length, and that retrieval gating can improve evidence use while its net benefit depends on preserving evidence recall. These results motivate evidence bottlenecks, alias-resistant representations, retrieve-then-reason architectures, verifier-mediated memory, and contrastive anti-poison training.

Retrieval-augmented generation (RAG) improves language models with retrieved evidence, but processing many long passages is costly and can introduce distracting information. Soft compression addresses this challenge by encoding passages as compact sequences of continuous memory embeddings before generation. However, existing methods typically assign each retained passage an identical number of memory embeddings, irrespective of its query-specific relevance. To address this, we propose AdaMem, a relevance-guided soft-compression framework that maps learned passage-relevance estimates to a query-dependent allocation of a fixed memory-token budget. A shared query-conditioned compressor produces both continuous passage memories and relevance scores in a single pass; a deterministic allocation rule assigns more memory tokens to higher-scoring passages and can omit low-scoring ones. Across six open-domain QA benchmarks, AdaMem consistently outperforms OSCAR (the closely matched soft-compression baseline that uses uniform allocation) as well as other soft-compression methods at matched memory budgets. Under standard 16$\times$ compression, AdaMem improves sub-string match by up to 3.2 points (5.5%) over uniform allocation baseline, with an average relative gain of 3.4%; under aggressive 64$\times$ compression the average relative gain grows to 14.6%, with a maximum of 9.8 points (19.7%) on PopQA. AdaMem matches the answer quality of the uncompressed at up to 4$\times$ lower inference latency than full context baseline. AdaMem retains an efficiency profile comparable to the uniform-compression baseline, while achieving up to $4\times$ lower inference latency than full-context inference. Thus, relevance-guided memory allocation is particularly effective when retrieval pools are large and the available memory budget is tight.

Cooking is a complex process that transforms raw ingredients into delicious and nutritious dishes, yet the recipes that encode this process remain largely free text; readable by people but not directly computable. Existing recipe collections capture fragments of this information, but no shared representation links a recipe's structured ingredient composition, its geo-cultural provenance, and its nutritional profile within a single queryable schema. We address this representation gap by formalizing a framework for recipe data structure that decomposes each recipe into typed ingredient entities, grounds those entities in a reference nutritional database, and annotates them with geo-cultural and dietary context. We present RecipeDB2, a structured compilation of 128,942 recipes with 35,474 ingredients from 32 regions and 99 countries. Ingredient phrases are parsed into seven culinary attributes using a transformer-based named-entity model; ingredients are linked to the USDA reference tables through a BERT embedding strategy (F1 = 87.90 on a manually adjudicated set of the 200 most frequent ingredients), yielding 148 nutritional parameters per mapped ingredient; a Random Forest classifier propagates 34 ingredient categories across the full vocabulary; and a deterministic, conservative rule set assigns each recipe a dietary style. Through RecipeDB2 (https://cosylab.iiitd.edu.in/recipedb2/), we demonstrate a scalable framework for making recipes computable, turning culinary heritage (long treated as an artistic rather than a quantitative object) into a data-driven analysis.

Speculative decoding accelerates language-model inference by letting a cheap drafter propose tokens that the target model verifies in parallel. Recent block drafters make drafting nearly free: a single backbone pass emits an entire block of draft tokens. Draft trees promise a further gain – several alternative continuations verified in one target forward – but existing constructions rank candidates by per-position marginals that ignore which parent a candidate extends, so on semi-autoregressive drafters wider trees mostly add mis-ranked nodes; and a tree of fixed size ignores how much speculation each decoding round, and each serving load, can support. We introduce TreeSpark, which reads a parent-conditioned distribution from the drafter's existing Markov head at negligible cost, calibrates it into an edge-acceptance estimate, and lets path survival govern everything else: best-first expansion, per-round stopping, and a load-adaptive serving policy. Sampling siblings without replacement, with matching residuals in recursive rejection, keeps decoding lossless at any temperature. Adaptive trees improve on matched fixed budgets at every temperature; against a tuned chain on the same drafter, TreeSpark accepts 15-25% more draft tokens per round and decodes 8-14% faster in single-request wall-clock, and under rising load it gracefully shrinks the tree back to the chain. Code and artifacts: https://github.com/PopSoda2002/TreeSpark

The evaluation of large language models (LLMs) on coding tasks has primarily focused on performance metrics such as pass@k. As LLMs continue to advance, many models now meet baseline performance requirements, reducing the discriminative power of performance-based evaluation alone. Yet a key question remains largely unexplored: how do LLMs differ in their coding behavior? We propose CLIC (Code Learning for Identification and Comparison), a visual analytics approach that characterizes LLM coding behavior through token-frequency analysis. CLIC represents each code sample as a feature vector of token frequencies and trains an interpretable decision tree to separate two LLMs' code sets. Beyond classification accuracy, we define two new metrics: robustness, which measures whether the two LLMs remain distinguishable as their most-discriminative tokens are progressively removed, and concentration, which measures whether the difference is driven by a few dominant tokens or spread across many. Interpreting numerous pairwise comparisons (across LLM pairs, tasks, and tokenization levels) and tracing the full analytical chain form an inherently multi-scale, hypothesis-driven exploration task. We therefore develop an interactive visual analytics system to navigate the comparison landscape, identify pairs of interest, and drill down into discriminative tokens and their code contexts. Case studies comparing 10 LLMs across 22 Kaggle ML tasks reveal actionable insights for LLM selection and prompt engineering.

Rubric-based evaluations of deep-research (DR) systems often obscure fine-grained factual failures in generated reports. We introduce CLAIMPROBE, a claim-level audit that decomposes DR reports into claims and measures hallucination, misattribution, citation hygiene, and necessary-fact recall against retrieved evidence. Using CLAIMPROBE, we find that strong DR pipelines can omit key evidence and misattribute claims even when their rubric scores remain stable. We then propose CLAIMWRITER, a hierarchical claim-based writer that extracts source facts, maps them to a query-derived outline, and drafts each section from a source-linked claim representation. Across three prior DR frameworks, replacing only the report writer with CLAIMWRITER reduces hallucination by 2.6 to 4.5 times and improves necessary-fact recall by 1.2 to 1.7 times, while largely preserving overall report quality. CLAIMWRITER also enables localized revision: when sources change, it propagates changed source facts into revised reports at the highest rate among update methods, while also being more cost-effective.

Large Vision-Language Models (LVLMs) have achieved remarkable performance across a wide range of tasks; however, they often inherit social biases from their training data, resulting in biased behavior when processing portraits from different social groups. Existing debiasing approaches typically compare token probabilities between the original and biased generations during decoding, but they are fundamentally limited by their reliance on a single, stereotyped viewpoint and fail to account for the diversity of social perspectives. Inspired by the social science principle that diversity fosters fairness, we propose Counterfactual Ensemble Decoding (CED), a novel framework that constructs multi-group counterfactual perspectives within the visual representation space and integrates them during decoding to promote equitable model behavior. CED first performs counterfactual steering in the visual space by identifying semantic directions associated with each social group and generating counterfactual representations along these directions, thereby offering diverse perspectives that disrupt stereotypical narratives. During decoding, CED locates the decoder layer exhibiting the greatest divergence among these perspectives and ensembles their token distributions using uncertainty-aware weights, prioritizing high-confidence tokens from different groups to yield a more balanced probability distribution that guides fairer generation. Extensive experiments on three social bias evaluation benchmarks demonstrate that \tool achieves substantial improvements over leading baselines, reducing bias by up to 47.97% across scenarios involving occupations, descriptors, and persona traits. Moreover, CED also preserves the core capabilities of the original model with minimal degradation.

Video understanding is a fundamental task for evaluating the capabilities of multimodal large language models (MLLMs). However, existing leading models have already achieved approximately 90% accuracy on the Video-MME leaderboard, suggesting that conventional single-turn video understanding tasks are becoming increasingly saturated and insufficient for assessing the intelligence of advanced MLLMs. Towards this end, we introduce VideoGAIA, an agentic video understanding benchmark for general artificial intelligence (AI) assistants. Moving beyond one-shot video question answering, VideoGAIA formulates video understanding as a multi-turn, tool-augmented interaction process, where models must iteratively perceive videos, invoke external tools, gather complementary information, and integrate multimodal evidence across turns. VideoGAIA contains 271 model-human co-designed tasks covering diverse and complex real-world scenarios. Each video-question-answer instance is independently verified by three human experts to ensure both correctness and appropriate difficulty. All evaluated MLLMs, including frontier models such as GPT-5.5 and Kimi-K3, achieve less than 60% accuracy on VideoGAIA, highlighting its value as a high-quality and timely benchmark for evaluating next-generation MLLMs. We hope that VideoGAIA will facilitate the transition from conventional video understanding toward agentic video understanding.

Long-running LLM agents act through tools, and a single step can send an email, merge a pull request, or wire a payment. The steering decision is the pre-commit choice at that boundary: proceed, or hold for human or policy review. We introduce SteerBench-Work, an incident-anchored, bidirectional benchmark for that decision in workplace agents across developer operations, customer service, finance, legal, medical, HR, and security. Release v2026-05 contains 106 scenarios anchored in public incidents, paired evidence-reversed mirrors, and calibration controls, with labels split nearly evenly between proceed and hold so the two error directions get near-identical numbers of chances. A model sees the proposed action and the available evidence, returns a gate decision, and is scored on whether it crosses or holds the boundary correctly. Across 30 model conditions the failures run almost entirely in one direction: models wrongly hold authorized, evidence-cleared work on 28.1% of opportunities and wrongly allow unsafe work on 1.0%. The hardest cases are risk-resolved commits, where signed or structured evidence has already cleared a real risk trigger, and models score markedly worse on evidence-reversed mirrors of famous incidents (63.8%) than on the incidents themselves (98.5%). General capability is not the same as steering calibration: higher-capability models often over-refuse at the commit boundary, and more reasoning can repair a weak gate while leaving a calibrated one flat. The public leaderboard is at steerbench.com.

Benchmark contamination is diagnosed with n-gram overlap, likelihood-based membership inference, or canary strings, and each needs something usually unavailable: the training corpus, a well-chosen test statistic, or foresight at release. A recent alternative reads it off a linear probe on internal activations. We show the natural way to do this does not work, specify one that survives measurement, then find that the correction making it work carries more variance than the null it is tested against. The protocol reports a zero-sum contrast on the depth profile of probe accuracy, recentred on a level-matched placebo baseline, tested against a label-permutation null, with the reference set twice the size of the suspect set. Each choice replaces a simpler alternative we rejected on measurement. Reporting the level of excess separability rather than its shape makes the false positive rate track the size of the analyst's own control set, 0.03 to 0.99 under a true null. Contrasting against a flat depth profile rejects a true null 0.72 of the time when surface decodability rises with depth, and loses all power when it falls. On real transformers the protocol fails a test the simulations did not pose. The recentring subtracts an estimate, and the permutation null holds it fixed. Re-estimated across split seeds on four audits of contaminated checkpoints, its standard deviation is 1.30 to 1.56 times the null's own in every arm: what is subtracted to remove a bias is more variable than what it corrects. The one nominally significant result, p = 0.0075, becomes 0.0745 once that variance is propagated, and no verdict is issued. The simulations missed this because their surface key is the covariate driving item variation; on real text it is a proxy, and degrading key quality in simulation reproduces it. We add a companion measurement and a widened null. No arm shows contamination.

While large language models can generate entire novels, there is little information about the level of formal variation in their output over many generations. Rather than asking whether individual passages can be identified as AI-generated, this study asks whether repeated AI generation can produce the same range of diversity which is found across human corpora. This paper contrasts six corpora based on generation source and target style: twenty novels generated using GPT-5.5 Thinking in a nineteenth-century British realist style, twenty novels generated using Qwen3-14B in a nineteenth-century British realist style, twenty novels generated using each of these models in a contemporary zero style, 205 nineteenth-century human-written British novels, and sixty-five contemporary human-written Zero-Style novels. At the document level, the research includes MATTR-500, Shannon entropy, average sentence length, readability, and punctuation rate measurements. The most robust and reliable result is compression of sentence structure. Repeated generations produce novels that vary far less from one another in sentence structure than human novels do. Compression is also present in the measures of readability, punctuation, and sentence length variability within novels. Lexical measures tend to be similarly compressed, with the exception of Qwen Zero-Style MATTR. Despite having distinct mean stylistic profiles, GPT and Qwen lack a stable pattern of cross-measure correlation. This article therefore distinguishes between variance overclosure, which represents a limited formal range between novels, and a more specific phenomenon of correlational overclosure. This means that an individual AI-generated novel may resemble human fiction stylistically, while a collection of AI-generated novels occupies a much narrower formal range.

Long-horizon egocentric memory transforms continuous first-person video and audio into a searchable record of past experiences. We demonstrate two bottlenecks in existing systems: indices built from context-poor captions are unreliable for agentic search, while retrieval ignores a question's temporal intent. To address both bottlenecks, we introduce EgoCITE (Egocentric Context-augmented Indexing and Time-aware Evidence retrieval), a long-horizon agentic memory framework for egocentric QA. EgoCITE comprises three components. EgoScheme uses local multimodal context to turn fragmentary video captions and speech transcripts into self-contained atomic memory indices. EgoIndex organizes complementary action, activity, utterance, and conversation representations into searchable multi-view memory indices at multiple granularities. EgoRetrv combines semantic search with question-conditioned temporal relevance scoring and curation of retrieved evidence. We evaluate EgoCITE on EgoLifeQA, EgoMem, and EgoR1-Bench in terms of answer accuracy and target-event retrieval alignment. EgoCITE improves accuracy over agentic memory baselines by at least 4.4–14.2% while achieving 36$\times$ lower cost than long-context LLM agents.

Strategic reasoning in Large Language Models (LLMs) within long-horizon environments is often limited by inconsistent subgoals. In these settings, finite attention resources prevent the model from maintaining strategic coherence over thousands of steps. This limitation leads to strategic drift, where localized decisions fail to sustain a coherent trajectory across reasoning. To address this, we introduce EpicStar, a framework that enables agents to learn memory as policy to tackle long-horizon reasoning. Specifically, the agent maintains a bank of successful past episodes as a heuristic alongside a working memory to track short-term environmental changes. During inference, a dynamic gating mechanism determines whether to execute a retrieved action directly or to perform new reasoning through a contextual fusion of the retrieved episodes and current working memory. Utilizing StarCraft II as the testbed, we evaluated EpicStar against diverse opponent styles. It significantly outperforms baseline methods, achieving higher win rates while consuming an order of magnitude fewer tokens, and it maintains this advantage consistently across difficulty levels and opponent strategies. Our findings provide compelling evidence that structured cross-episode memory is essential for enabling LLM agents to perform robust, long-term strategic execution in dynamic, autonomous settings.

Language model classifiers with explanations are used for moderation, routing, topic triage, and low-resource annotation. We study black-box auditing when the defender has only clean calibration data without trigger information but can ask the classifier for a label plus a short rationale or quoted evidence. We introduce Groundedness Drift, a lightweight score measuring whether the answer summary remains grounded in the input. Across two 7B backbones, five datasets, and four common non-adaptive OpenBackdoor-style attack families, Groundedness Drift achieves higher AUROC and lower residual target ASR than every compared detector in all cases at a nominal 5% clean-FPR budget. We then evaluate Unsupported Groundedness, a multi-probe escalation for explanation-camouflage stress cases. Unsupported Groundedness improves signals but does not close the adaptive gap.

Intensional operators are often treated as quantifiers over possible worlds, parallel to the treatment of determiners as quantifiers over individuals. Yet individuals introduced in intensional contexts cannot serve as antecedents to later pronouns as easily as those introduced in merely quantificational contexts. For instance, "Everyone is eating a cheeseburger" may be followed by "They are large", where "they" refers to the cheeseburgers being eaten. However, as Stone (1999) points out, the similar "Andrea might be eating a cheeseburger" does not support later anaphoric references such as "It is large" or "They are large". Stone (1999), Stone and Hardt (1999), and Brasoveanu (2010) address this by requiring a pronoun's value (its referents) to exist in the world of evaluation, ruling out anaphora from non-veridical intensional contexts. We show, however, both cases where such anaphora is disallowed even when the pronoun's referents clearly exist and cases where it is allowed even though they might not exist. We argue that intensional anaphora is best captured using a description-based rather than value-based account. A pronoun presupposes that its corresponding antecedent description is instantiated in each world of the context set. Thus, there must be a cheeseburger being eaten by Andrea in every candidate world for "It is large" to be felicitous after "Andrea might be eating a cheeseburger". We implement our proposal via a new logic, building on Keshet (2018) and Abney and Keshet (2022), called Plural Intensional Presuppositional predicate calculus (PIP). Each PIP formula translates directly into standard first-order predicate calculus with set abstraction, providing a classical foundation for this work.

In 2023, a New York judge sanctioned two attorneys in Mata v. Avianca for filing a brief with hallucinated citations generated by ChatGPT. Such failures are largely caught by database lookups; the harder problem is detecting citations that point to real cases but do not support the propositions for which they are offered – a failure mode that existing evaluations of LLMs for legal use cases largely overlook. In this paper, we study proposition-level citation support verification through controlled perturbations of real legal citations obtained from two legal corpora, either replacing the cited case or changing only the pinpoint page within the same case. We evaluate fourteen model configurations on the resulting examples. Models catch 93-100% of wrong-case corruptions. They catch only 37-61% of wrong-pinpoint corruptions on court opinions and 52-83% on legal briefs. When models fail to catch wrong-pinpoint corruptions, they accept the citation based on topical overlap rather than page-level support. Scale and extended reasoning narrow the gap but do not close it: GPT-5.4 with high reasoning effort still misses 40% of pinpoint mismatches on court opinions and 18% on briefs. Prompting the model to verify support at the cited page improves recall, but it also raises the false positive rate. Recognizing the right legal topic and verifying support for the cited proposition are distinct capabilities, and current models conflate them.

Privacy documents (e.g., privacy policies) are a central mechanism through which digital services disclose data practices and seek user consent. Over the past decades, research on privacy documents has expanded significantly, encompassing not only traditional privacy policies but also short notices (e.g., privacy labels) and interface-level transparency mechanisms. As this research area continues to grow, it has become increasingly difficult to obtain a coherent view of how privacy documents are created, analyzed, evaluated, and maintained across their lifecycle. This SoK provides a unified, lifecycle-oriented view of privacy documents from a software engineering perspective. We systematically review and analyze 290 papers published between 2010 and 2025, organizing them around five research questions that examine how privacy documents are (1) defined and scoped, (2) generated, (3) analyzed and extracted, (4) checked for inconsistencies and noncompliance, and (5) evaluated and improved for usability. Building on our findings, we identify 15 key research trends and 21 open opportunities. We further chart four broader research directions that highlight (i) emerging challenges in AI-centric platforms, (ii) the need for diverse and up-to-date data foundations, (iii) LLM-based unified policy-code analysis, and (iv) dual usability for end-users and developers. We hope this SoK provides a shared foundation for future research on privacy policies and privacy documents.

Large language models (LLMs) cannot retain post-deployment experience without parameter updates. We introduce DIVE, a diversity-driven framework that enables frozen LLMs to improve by evolving persistent natural-language skills from task experience and verifier feedback. These skills encode reusable reasoning procedures, verification strategies, common failure modes, and output constraints and are both executed and revised by the same underlying model without access to a teacher model. Since natural-language skill evolution is a stochastic, non-convex search process, optimizing a single skill trajectory can overfit to sampled experience or converge to a suboptimal solution. DIVE mitigates this optimization variance by independently evolving multiple skill populations from bootstrapped experience, adaptively refining them through diverse transformations, and jointly selecting a complementary set of skills. Across six mathematical and logical reasoning tasks and multiple model families, DIVE consistently outperforms existing reasoning methods, prompt-optimization approaches, skill-development frameworks, and memory-based baselines. It achieves rapid self-improvement from accumulated experience, obtaining substantially larger performance gains with fewer rollouts than parameter-based methods such as SFT and GRPO, and prompt optimization with GEPA. Further, the resulting skills transfer across model scales and families, enabling smaller models such as GPT-5-nano to match or outperform larger counterparts, i.e., GPT-5, under conventional prompting. These results establish diversity-driven skill evolution as an effective, interpretable, and parameter-free approach to LLM self-improvement.

Trained transformer models develop privileged bases: coordinate axes whose statistics differ from the rest of the residual stream. But what kind of direction does such a basis select? We investigate the prediction direction, the unembedding direction of the token a model currently predicts, and find that it functions as a content-defined privileged anchor. Measured with respect to this anchor, residual-stream variation is geometrically and behaviorally stratified by proximity to the prediction. The stratification holds in all eighteen models tested (dense and mixture-of-experts, 7B-120B, base and instruction-tuned). A narrow, scale-invariant prediction interface concentrates readout-relevant structure, while the vast prediction-distal complement expands with model scale. Because the prediction direction sits nearly orthogonal to the principal variance axes, variance-based analyses recover this organization only partly, and the shortfall grows with prompt heterogeneity. Anchoring reveals a steep geometric gradient: prediction-proximal regions are highly structured and cluster related prompts, while the complement is flatter and anti-discriminates among prompt groups. The interface is a narrow slice but functionally decisive. Disrupting the variance directions closest to the prediction causes immediate divergence and frequent task-frame shifts; disrupting the next level down delays divergence and preserves framing. The complement is weakly readout-aligned per direction yet causally and temporally load-bearing, and behavior is driven by direction rather than magnitude. These results establish the prediction direction as a privileged anchor distinct from previously described coordinate axes, and give a geometric account of how high-dimensional computation coexists with linear readout.

Large language models are increasingly deployed in settings that require simultaneous adherence to multiple explicit constraints - reasoning structure, safety boundaries, output schemas. Individual constraints are handled proficiently, but the compositional regime, where many must hold jointly, remains poorly characterized: how rapidly does performance degrade, what governs the degradation, and can the collapse be mitigated? We introduce Constraint Saturation Evaluation (CSE), a procedurally generated benchmark that systematically varies the number of simultaneous constraints (k), with every constraint scored by a deterministic, rule-based verifier and zero LLM-judge involvement: 15 models, 36 constraint types, 369,753 checks at k=1-12. Three findings emerge. First, per-constraint pass rate decays gradually and predictably, while the chance of satisfying all k constraints collapses - a model passing individual constraints at ~41% at k=8 succeeds on all eight just 5.7% of the time. Second, constraints do not degrade equally: structural constraints lose 2x more baseline capability per added constraint than lexical ones, ordered by a comprehension-maintenance gap that separates constraints requiring sustained tracking from binary decisions immune to composition. Third, failures are nearly independent, which is what makes the accumulation multiplicative; the residual coupling that does exist tracks shared output features rather than pairwise interference - a wrong sentence count fails every constraint that reads it. Reliable instruction following breaks down beyond 5-6 simultaneous constraints: probe-level success falls below 50% at 7 constraints for the strongest model, and at 3 or fewer for 12 of 15.