Papers

Unsupervised feature selection seeks a compact subset of informative features without access to class labels, making feature utility difficult to define. Existing UFS methods therefore rely on indirect structural criteria, such as similarity preservation, locality, sparsity, cluster geometry, or reconstruction quality. In this paper, we instead study UFS through representation consistency and propose Inverted Contrastive Learning for Unsupervised Feature Selection (ICLFS), a feature-wise contrastive framework that reformulates UFS as a representation learning problem over features rather than samples. ICLFS first inverts the data matrix so that each feature is represented by its sample-profile vector, then constructs multiple masked positive views together with a shuffled negative view, and learns projector-space representations that remain consistent across these structured perturbations under an InfoNCE-based objective. Motivated by recent findings that cosine-based and InfoNCE-based training affect embedding norms, we use projector-space embedding magnitude as the saliency signal for ranking features. The resulting norm-based ranking is subsequently refined through Laplacian-Gated Ranking Correction, which suppresses locally redundant candidates while preserving salient ones. Extensive experiments on 12 benchmark datasets show that ICLFS achieves the best clustering accuracy on 10 datasets against both classical and neural baselines under the standard clustering-based UFS evaluation protocol, while remaining competitive on the other two. These results show that feature-wise contrastive representation consistency provides a strong and effective alternative to neighborhood, cluster, and reconstruction-based UFS formulations.

Large language models (LLMs) are increasingly trained to decline queries that fall outside their knowledge (knowledge-based refusal, KR) or violate safety policies (safety-based refusal, SR). Although KR and SR result in superficially similar responses, they have largely been studied in isolation, leaving open whether they share an underlying mechanism. We address this gap with a systematic study on a new dataset of 213 contrastive quadruples that jointly probe both refusal types. We find that KR and SR are governed by overlapping yet distinguishable mechanisms. Both share a refusal direction, yet the overlap is asymmetric: SR signals transfer more strongly to KR than the reverse. Type-specific specialization emerges mainly in upper layers, with KR aligning with uncertainty- and knowledge-related representations and SR with safety- and policy-related ones. We thus characterize refusal as a commit-then-specify process: a shared initial mechanism commits to refusing, then type-specific features in later layers specify whether the grounds are epistemic or normative.

Large language models (LLMs) increasingly handle in-context learning (ICL) tasks where a long, novel context defines the rules, knowledge, and output schema for a series of questions. On benchmarks that grade against every detail of the context, even strong open-weights models pass only 12-16% of tasks: a single overlooked rule fails the whole response. We argue this brittleness is structural: the dominant "read-and-reason" paradigm asks the model to extract, plan, generate, and self-verify in one forward pass. We therefore ask whether explicit context compilation can fix it, how it compares to existing long-context strategies (gist retrieval, multi-agent self-play), and where the resulting harness benefit holds across task structure and model scale. We propose the Context Compilation Architecture (CCA), whose central novelty is a typed intermediate representation (IR) with fixed slots (rules.{must_do, must_not, conditional}, output_spec, available_tools, data_profile) into which any prose context is compiled once; executable verifiers and a violation-gated correction loop follow as downstream consequences. On CL-bench (1,899 tasks across 4 open base models), CCA outperforms vanilla prompting and two long-context baselines (ReadAgent-P, Ctx2Skill) on every base model, lifting Kimi K2.5 from 15.4% to 21.4% with gains concentrated on rule-dense sub-categories. Code and cached completions are available at https://github.com/TonyQJH/cca-emnlp2026.

The Mutual Reinforcement Effect (MRE) asks whether a fine, span-level and a coarse, document-level task help each other when one model handles both. We test it in multimodal document understanding on three corpora, two of receipts and one of scanned business forms, comparing single-task, joint and conditioned training, which puts one granularity's gold output in the other's prompt during training only. We build Doc-MRE, an annotation layer pairing gold field extraction (point) with four document-level facets (line), from a three-judge LLM committee under a pre-registration, validated by blind re-annotation. One predicate, fixed in advance: at a shared recipe, a regime reinforces if it beats the matched single-task model on both granularities. Mixed joint training, the arrangement prior MRE work assumes, reinforces on no corpus at the main scale: it is below both single-task models on CORD and trades one granularity for the other on the two others, as single-task tuning does. Conditioned training reinforces on two of the three, CORD (+0.5 point, +4.8 line) and the forms corpus (+7.2 point, +11.0 line), resolvably on the coarse side and directionally on the fine one, and trades on WildReceipt; at that recipe no alternative measurably beats it on either side anywhere. Two byte-identical-prompt controls separate content from format: shuffled conditioning destroys the coarse-side skill but costs the fine side far less, and a neutral-content control reproduces the whole fine-side gain on WildReceipt, which is therefore prompt structure but buys nothing resolvable on the other two. On the forms corpus conditioning buys collapse avoidance: mixed training and the neutral control both assign the majority semantic label to all 50 test documents; only conditioning recovers the gold distribution. Probes find the information decodable under every regime with no resolvable increase under conditioning.

When contextual information conflicts with knowledge stored in model parameters, activation directions can be used to decode and steer which source the model follows. However, successful steering does not establish that the unedited model uses those directions to choose between sources, or that they remain effective across tasks. To test these possibilities, we vary the stated authority of contextual claims while holding their content fixed. We first estimate authority directions from prompts in which context and parametric knowledge agree, then test their causal contribution when the two sources conflict. Interchanging naturally occurring activation values along these directions between matched high- and low-authority prompts reproduces 30–68% of the authority-induced shift in source choice across Qwen, Llama, and OLMo models, whereas matched controls reproduce almost none. We next ask what transfers across tasks: the learned direction versus the activation values exchanged along it. Using a direction learned on another task closed 9% of the source-choice gap, compared with 57% when learned on the task being evaluated. Both interventions exchanged activation values from the evaluated task. In a separate experiment, we kept its learned direction but exchanged values taken from another task, which closed 68% of the gap. These results show that authority-related activation values can causally influence source choice across tasks when inserted along directions learned for the task being evaluated.

A transformer carries each token's state from layer to layer, and the whole vocabulary carried together forms a cloud that moves with depth. We ask whether a trained network moves this cloud the way optimal transport would: at the cheapest cost, and along the map that pairs each token with its optimal destination. We measure both on Pythia-160m and Pythia-410m, with an exact assignment between consecutive layer clouds, a measured sampling floor, calibration on couplings known to be optimal, and a split of the cost into the common shift of the cloud and the token-specific moves. At the last layer, both models move their tokens where the optimal-transport map sends them, at the optimal cost for Pythia-410m and slightly above it for Pythia-160m. At the first layer they do not. In between, single layers can be judged on cost at only two of ten transitions, and blocks of several layers move the cloud at close to the optimal cost. The agreement at the last layer is much weaker at initialisation (0.64 against 0.86) and grows with training.

Large language models increasingly generate research ideas, yet judging their novelty or feasibility at generation time does not establish whether they anticipate subsequent work. We introduce IdeaForecastBench to evaluate research idea forecasting. Given a community's literature up to a cutoff, a system produces up to five ranked ideas, which are evaluated against later papers. The benchmark comprises 624 rolling episodes across 52 topics, with a fixed retrieve-then-judge protocol and separately reported results from two judges. We compare five history-compression strategies across GPT-4.1, Qwen2.5-7B/14B, and Qwen3.5-9B, together with a learned Mode-Decomposition Forecaster (MDF). Under the primary GPT-4.1-mini judge, Summary improves on Direct in Hit@5 and Precision@5 across all four backbones. Qwen2.5 scores above GPT-4.1, whereas Qwen3.5 scores below it. An outcome-blind assessment finds that Qwen2.5 produces broader forecasts, but does not identify how much breadth contributes to its advantage. Threshold and judge diagnostics further clarify the limits of interpreting realization as precise anticipation. IdeaForecastBench provides a common task for studying which research ideas a community subsequently pursues and how reliably this outcome can be measured.

Large language model (LLM)-based multi-agent systems (MAS) have shown strong potential for solving complex tasks, yet their development forces a tradeoff: code frameworks are expressive but engineering-intensive, while no-code builders simplify authoring but constrain agent interactions to author-defined workflows. We present ChatDev 2.0: DevAll (hereafter DevAll), a no-code platform for building, executing, and inspecting heterogeneous MAS that delivers both high expressiveness and ease of use. In terms of expressiveness, DevAll pairs a declarative executable graph abstraction with a cycle-aware execution engine, so that heterogeneous agents and dynamic and cyclic interactions can be represented and executed within a single framework. For ease of use, an integrated visual interface lets users author, run, monitor, and inspect MAS, including human-in-the-loop steps, entirely without writing code. Experiments demonstrate that DevAll reproduces state-of-the-art MAS across three representative tasks at competitive performance and without task-specific orchestration code, highlighting its effectiveness as a general-purpose platform for LLM-based MAS. DevAll is available at https://github.com/OpenBMB/ChatDev.

Large Vision-Language Models produce fluent image descriptions but offer limited semantic control: users cannot reliably specify whether captions should emphasize attributes, relations, or particular image regions. We present Fine-grained Captioning Control Using Scene Rewards (FoCUS), a controllable image captioning method that lets users steer captions toward specific semantic emphases through natural-language control prompts. The core idea is a prompt-conditioned control objective based on scene-graph-aligned component scores. Generated captions are parsed and aligned to scene-graph components such as objects, attributes, and relations. These components are differentially weighted, including negative weights, according to the requested emphasis. We optimize this objective with GRPO and further improve its reliability through a stricter object validity threshold and reasoning-based verification for attribute and relation scoring. To evaluate controllability, we introduce Semantic Control and Precision Evaluation (SCoPE), a benchmark with contrastive Include/Avoid constraints for measuring both target content coverage and out-of-scope suppression. Experiments on two VLM backbones show that FoCUS consistently improves controllability and fine-grained caption quality without degrading general caption performance.

The SAIR Mathematics Distillation Challenge on Equational Theories asks a solver to classify whether one magma identity implies another and, for either verdict, to return a certificate accepted by a deterministic Lean judge. We present a single-file solver organized as a cheapest-first cascade. Its false branch combines coefficient tests over structured algebra families, bounded finite-model search, an explicit central-groupoid witness, and several infinite-carrier witnesses. Its true branch is a proof-producing ordered unit superposition procedure with Knuth-Bendix ordering, bidirectional demodulation, indexing, memoised substitution, and anytime size deepening. Search results remain outside the trusted base: successful derivations are replayed as small Lean terms, and countermodels are rechecked by the competition judge. The frozen solver is a 189,504-byte Python file with SHA-256 f2392533c9f4c03b.... In local runs through official judge revision 2848228, it produced accepted certificates for all 1,889 rows of the six public sets with no language-model calls. Separate measurements recorded full agreement on the 800 published Stage 1 evaluation-distribution problems, 100 accepted rows in the canonical Marathon manifest without tokens, and 200 accepted rows in the hosted playground. These are regression and playground measurements, not a leaderboard result and not evidence about a hidden set. All quantitative claims are tied to immutable result ledgers; the paper makes no completeness or comparative-superiority claim.

LLMs have been rapidly adopted across writing tasks, prompting the development of tools for detecting LLM-generated text. Yet, these tools largely measure how much of a document's surface text was written by an LLM and aren't fundamentally designed to measure how much of the information content or ideas originated from the LLM itself rather than being supplied by the user in the prompt. In this work, we design a framework that measures how much value a person adds on top of what a language model could have easily produced by itself. The method requires no training or labeled data and never scores the document's surface text, insulating it from stylistic confounders. Instead, it extracts the document's content at increasing levels of granularity, uses an LLM to reconstruct the document from each partial representation, and compares these reconstructions with those produced from the task description alone. We call this framework Value Over Language Model (VOLM), which measures a document's contribution relative to a replacement-level document that an LLM could produce from the task description alone. We evaluate VOLM with a specific instantiation of this framework across three domains: news articles, ICLR peer reviews, and argumentative essays. VOLM separates human-authored documents from matched LLM-generated documents produced from generic task descriptions, while remaining substantially invariant to content-preserving transformations, including LLM-based reconstruction and round-trip translation. We further find that increasingly constrained content extractors reduce residual differences between LLM-generated and humanized text, demonstrating the importance of disentangling informational content from stylistic variation. We hope these results encourage further work on specialized instantiations of the framework and on assessing human contributions in LLM-assisted writing more generally.

Retrieval-Augmented Generation (RAG) is highly sensitive to retrieval noise: when retrieved documents mix informative and irrelevant context, LLMs are easily distracted, leading to hallucinations. To overcome this, we propose SCoNE (Selective Context-aware Neuron Editing), a training-free model editing approach that improves retrieval noise robustness by selectively strengthening context-aware FFN neurons that are identified by both high attribution and high cross-input variability. SCoNE requires only a small number of mining samples, no fine-tuning, and no inference-time overhead. Across various knowledge-intensive question-answering benchmarks and two LLM backbones, SCoNE consistently outperforms competitive baseline methods. Our code is available at https://github.com/HYU-ARK-Lab/SCoNE.

Article-level news stance detection aims to identify the perspective of news articles toward social issues. Despite advances in stance detection and its importance for trustworthy media environments, news articles pose distinct challenges because their stances are often implicit, subtly conveyed through journalistic framing, and embedded in long, structurally complex texts. To address these challenges, we introduce VFStance, which leverages visual framing to make implicit stance cues more explicit via image generation. In evaluation experiments, we demonstrate the effectiveness of VFStance over existing methods and the contribution of visual framing to its performance. Finally, a controlled user study (N=200) in a snippet-based news consumption setting further demonstrates that VFStance can make stance signals visually salient and highlights its potential use beyond automated stance detection.

Creative generation tasks, such as narrative writing and scientific ideation, demand both high-quality outputs and distinct responses across independent runs to maximize exploration. Multi-Agent Debate (MAD) has shown strong quality gains on factual and reasoning tasks, making it a natural candidate for creative generation. However, we find its convergence-driven design actively suppresses output diversity across independent runs, creating an inherent trade-off with creative tasks. We theoretically show that preserving diversity among agents within each debate session is a necessary condition for achieving diverse outputs across independent runs. Building on this finding, we propose Creative-MAD, which introduces two synergistic interventions to sustain agent divergence. Specifically, Cognitive Lens Assignment counters identity drift by anchoring each agent to a distinct and persistent cognitive mode, while Embedding-based Peer Selection counters majority pull by limiting each agent's context to its most semantically distant peers. Experiments across four creative benchmarks demonstrate that Creative-MAD significantly enhances both lexical and semantic diversity while maintaining MAD's output quality.

We describe the SciTrue team's participation in both subtasks of the NTCIR-19 SciClaimEval task~\cite{sciclaimeval}, which asks systems to verify scientific claims against the tables and figures of a paper. Rather than tuning a single model, we benchmark eleven frontier and open multimodal models under one honest, per-sample protocol and combine them with light, transparent post-processing. On the official, blind test leaderboard (Section~\ref{sec:results}), SciTrue placed first by a clear margin in three of the four evidence-category/subtask combinations, and tied for first on the primary metric in the fourth. Three findings explain the result. First, strong instruction-tuned models are already competitive: Claude Opus~4.8 and Gemma-4-31B each exceed the strongest public baseline (o4-mini), and GPT-5.5 and Claude Fable~5 lead both subtasks (97.7 on Subtask~2). Second, the task's pairing structure is the largest lever: a leak-free pair prior that recovers the Supported/Refuted pairing from the claim text alone (a visible field) and assigns Supported to the higher-confidence evidence raises Subtask-1 pair-accuracy from 72.2 to 93.5, far more than any model swap or ensemble weighting. Third, a case-by-case audit finds that most residual errors are visually-undetectable label-mapping swaps or dataset label noise, so measured accuracy understates the true ability and the fixable-by-modeling headroom is small. Controlled fine-tuning, distillation, and agentic consistency-checking support the same conclusions, and we document throughout a measurement leak—label information reaching a system through the packaging of the data rather than its content—in which the released file ordering encodes the label, including one instance that briefly misled our own pipeline.

Retrieval is the first stage of modern search and advertising systems, selecting a candidate set from a large item universe for downstream ranking and auction. Recent work increasingly leverages LLMs to improve retrieval through query expansion, data synthesis, and retrieval-feedback training. However, the generative component is typically used for query-side augmentation, while final matching is still delegated to a downstream retriever. We introduce CoGR, a retrieval framework that instead trains LLMs to directly construct retrieval representations on both query and item sides. Each generator produces a compact set of keywords, which are matched directly through an inverted index, preserving compatibility with existing keyword-based retrieval infrastructure. CoGR uses a two-stage training pipeline. Supervised fine-tuning first establishes an aligned keyword space, after which co-evolving reinforcement learning alternately optimizes the query- and item-side generators with GRPO against the opposite side's frozen index. Both sides optimize the same query-to-item retrieval $F_1$ objective: the query side receives retrieval $F_1$ directly, while the item side receives a counterfactual marginal reward measuring the change in query-side $F_1$ caused by its generated keywords. Across 10 representative sparse, dense, and generative baselines, CoGR achieves the best performance on both an internal APP Marketplace dataset and the public WANDS benchmark, improving $F_1$ over the strongest baseline by $10.9\%$ and $36.1\%$, respectively. Further analysis shows stable co-evolution and increasingly aligned query–item keyword spaces over training.

Large Vision-Language Models (LVLMs) achieve strong performance on image-grounded text generation and visual question answering. However, it remains difficult for them to comprehensively and accurately describe the factual relations among the entities and concepts associated with the objects depicted in an image. In this work, we propose a framework that efficiently exploits factual information from a knowledge graph via retrieval-augmented generation (RAG), with the goal of enabling LVLMs to generate detailed and accurate image explanations. Specifically, our method alternates between answer generation and knowledge-graph retrieval, and controls the search using a correctness judgment, thereby acquiring the necessary and sufficient factual information efficiently. We also construct a knowledge graph for the artwork domain (ExpArt-KG), in which the correspondence between images and entities is unambiguous. Applying the proposed method to this knowledge graph, we show experimentally that it improves the level of detail of artwork explanations and reduces the retrieval cost of external knowledge while maintaining generation quality comparable to that of iterating a fixed number of times.

A prominent paradigm in inference-time alignment employs lightweight supervisors to steer Large Language Models (LLMs). Through empirical analysis, we identify a structural mismatch in this paradigm: weak supervisors exhibit pervasive high entropy across the vast majority of tokens, yet prevailing dense intervention approaches mandate supervision at every decoding step. This leads to frequent low-confidence interventions that can disrupt valid base-model reasoning and incur substantial utility costs. To resolve this, we propose TUSA (Trust-based Uncertainty Sparse Alignment). Moving away from continuous oversight, TUSA reframes alignment as a dynamic arbitration process, introducing an uncertainty-aware arbiter that authorizes intervention only when two conditions are met: the supervisor is confident and the token is semantically salient. This mechanism effectively filters out uncertainty-driven noise and redundant supervision. Extensive experiments across multiple models and benchmarks show that TUSA consistently improves both safety alignment and general helpfulness. By bypassing approximately 50% of alignment steps, it not only enhances safety preference by up to 15.6%, but also boosts general preference rates by up to 12.0% compared to the dense baseline, demonstrating that selective, high-precision alignment can outperform continuous supervision.

Prompt optimization can improve multi-agent LLM systems, but the prompts being optimized often serve two entangled roles: generating task-relevant content and specifying execution-critical protocols, such as message routing, output formatting, and termination signals, on which the underlying code relies. As a result, a prompt edit intended to improve content generation can inadvertently corrupt the protocol and cause the entire agent pipeline to fail. Our key observation is that these two roles have different representations: execution protocols are typically structured, while task-relevant content is usually expressed in unstructured language. Based on this, we propose control-data flow separation, where execution-critical control is represented as typed, validated program objects, while task-relevant language remains the optimizable data flow for agent communication. This design allows optimizers to improve multi-agent behavior without exposing the routing or formatting interface to prompt drift. Across synthetic reasoning, collaborative review generation, and insurance rating workflows, our framework empirically achieves 100% eventual protocol validity while consistently improving task performance.

LLM-based user simulators are increasingly used to evaluate autonomous agents at scale, in place of costly human evaluations. Despite this promise, these simulators exhibit "assistant bias," a tendency to cooperate and pursue task goals. They rarely reproduce the frustration or disengagement that real users exhibit, compromising evaluation validity. Prior work outlines that this bias is baked in during model training, which role-playing prompts fail to override. We analyze this bias from model activations, extracting a user role vector by contrasting how the model represents user versus assistant perspectives on the same dialogue. We observe two findings: (i) the user direction is identifiable in activations, elicits user-like behaviors, and captures characteristics distinct from assistant traits; and (ii) although user-role activation associates with simulation realism and steering strengthens it, it can exaggerate user behaviors and override individual user profiles. Together, our findings provide a representation-level analysis of LLM user simulators, confirming that assistant bias is structurally identifiable and that user behavior can be directionally analyzed.

Machine unlearning for large language models (LLMs) often assumes that a pre-defined forget set matches what the model has memorized, but this frequently breaks in realistic privacy settings where the original training data is inaccessible. We term this gap forget-set misalignment and identify two cases. In Under Unlearning, the forget set omits memorized information and leakage persists. In Out-of-Knowledge Unlearning, the algorithm is driven to "forget" knowledge the model never learned, perturbing parameters and degrading utility. Using gradient-level analysis, we show these behaviors arise from misaligned unlearning targets rather than specific optimization choices. We then propose CONfession-to-Forget-Set (CONFS), a data-blind framework that constructs model-aligned forget sets by eliciting and formalizing the model's memorized knowledge. Across synthetic, multimodal, and real-world benchmarks, CONFS approaches Gold-standard performance on several metrics and achieves a competitive forgetting-utility balance, while preserving utility better than other data-blind forget-set constructions.

Reranking methods, such as Minimum Bayes Risk (MBR) decoding and Quality Estimation (QE) reranking, have been widely used in modern neural machine translation (NMT) to select an output from a set of candidate hypotheses. However, the performance gains come at the cost of high inference latency. Existing acceleration methods target MBR decoding and reduce only the reranking computation, leaving QE reranking unaddressed and candidate generation—which can be the larger computational bottleneck—largely untouched. In this work, we propose Quit (Quantifying Uncertainty for Incremental Termination), a novel early-stopping strategy for the entire generation–reranking pipeline. Quit treats candidate generation as a sequential decision-making process under uncertainty. It incrementally generates and reranks candidates, stopping when the best reranking score stabilizes. Comprehensive experiments with three NMT models across 19 language pairs show that Quit achieves end-to-end speedups of $1.47$–$2.66\times$ for MBR decoding and $3.43$–$4.12\times$ for QE reranking while preserving automatic metric scores.

Ensuring factuality remains a critical challenge for deploying LLMs in high-stakes settings. Existing hallucination detectors usually operate at a single level: claim-level methods provide interpretable factual units, while span-level methods localize unsupported text. Bridging these views is costly, as LLM-heavy pipelines require multiple decomposition and verification calls, and modular systems need additional claim-to-span alignment. We propose Enoki, an Open Information Extraction framework for multi-level hallucination detection. Enoki extracts text-anchored relational facts, verifies them against evidence, and projects unsupported facts back to hallucinated spans. This shared representation enables claim-level verification and span-level localization without requiring separate alignment. Enoki supports LLM-based, encoder-based, and rule-based extraction regimes, balancing accuracy and inference cost through a common interface. Experiments show that Enoki remains competitive with strong claim-level systems while using fewer resources and achieves superior performance on fine-grained span- and entity-level localization. We also release EnokiQA, a dual-granularity dataset with aligned claim-level verification and span-level localization annotations.

Large language models (LLMs) have shown proficiency in various software engineering tasks, such as code generation and translation. However, a key limitation in their performance may be their (lack of) understanding of programming-language semantics. Even when explicit semantics are given, it remains unclear whether LLMs apply those rules or lean on priors learned during pre-training instead. We study if LLMs lean on priors or given semantics with a novel task–Program Executability Prediction (PrEx)–that asks models to predict whether a program is semantically valid or invalid (and, if invalid, which formal rule it violates) given the program's syntax and operational semantics. Because PrEx requires both valid and invalid programs, we build a dataset with systematically generated invalid transformations derived from valid programs. We evaluate open-source coding LLMs under two semantic formalisms and two semantic shifts across Human-Written, LLM-Translated, and Fuzzer-Generated program splits. Our findings show that LLMs lean on pre-training priors rather than systematically applying the given rules, performing especially poorly on modified semantics and degrading further as program complexity increases. PrEx is available at https://github.com/EngineeringSoftware/prex.

Item-sensitivity, defined as whether a model's choice depends on the specific input rather than on its own output prior, is widely reported as evidence of task competence. We show this evidence is necessary but not sufficient using a forced-choice signalling task abstracted from the board game Deception: Murder in Hong Kong. In this environment, the reference points against which a coordinate should be judged (a fit-maximising strategy, a posterior-maximising strategy, and uniform random selection) are all computable in closed form. Across seven language models, two model families, a post-training ablation, and three independent scoring rules, every one of 21 model-by-rule cells is reliably item-sensitive. Yet 8 of those 21 cells are not statistically distinguishable from a chooser that ignores the item and selects at random, and 5 score worse than random at describing the target. Item-sensitivity and distance from random correlate at only r = 0.30. We call this consistency without alignment and argue it generalises to any evaluation that relies on item-sensitivity, permutation consistency, or self-consistency without an independent reference for the measured quantity. We further find that a literal-similarity baseline with no pragmatics outperforms most tested language models, that adding a pragmatic layer over two baseline similarity sources moves choosers toward random rather than toward the Bayesian reference, and that a standard labelled multiple-choice format carries no measurable content signal here. All results represent the model side of a pre-registered instrument; a matched human condition is designed and piloted but not yet collected.

Cultural fine-tuning has become the de facto paradigm for building culture-aware large language models (LLMs), yet existing optimization exclusively for alignment scores provides an incomplete portrait of cultural fidelity by systematically obscuring inherent cultural diversity. This unidimensional evaluation lens prompts a fundamental question: do models genuinely perceive distinct cultural nuances, or do they merely memorize dominant cultural values? To address this, we propose a synergistic evaluation framework that jointly formalizes cultural alignment and diversity. Through extensive benchmarking of six mainstream LLMs on the World Values Survey, this framework uncovers a systematic and critical trade-off: the pursuit of cultural alignment consistently incurs an acute expense of diversity, leading to severe "cultural flattening." Investigating this behavioral shift, we demonstrate that these superficial alignment gains stem from models artificially anchoring to dominant majorities, converging onto a monolithic response pattern that wipes out the heterogeneous distributions inherent to human groups. Crucially, our mechanistic analysis suggests that this diversity collapse is not merely a behavioral anomaly but more likely a structural consequence of the low-rank bias inherent in neural network optimization. Therefore, our findings expose the limitations of current post-training paradigms and call for a shift toward alignment objectives that preserve cross-cultural pluralism.

Multimodal memory offers a scalable interface for long-video question answering, but existing methods often retrieve captions, frames, transcripts, summaries, or graph facts as isolated fragments. Although searchable, such fragments are not generation-ready: language models must reconstruct cross-modal and temporal alignments at inference time, when context is limited and attribution is difficult. We propose EM^2Mem, an event-centric multimodal memory framework that binds heterogeneous evidence to event anchors during memory construction. Each event-indexed memory cell aligns multimodal records, temporal context, graph-linked relations, semantic facts, and provenance, enabling compact evidence readout over grounded multimodal events rather than modality-specific fragments. Across three long-video QA benchmarks, EM^2Mem improves average accuracy over the strongest memory baseline by 2.0, 2.4, and 3.7 points, improves strict event-level Top-5 evidence recall by 7.0 points, and reduces per-query latency by 4.67 times and total inference tokens by 63.66% (The code will be integrated into https://github.com/zjunlp/LightMem).

Multimodal large language models (MLLMs) are increasingly provided with contextual evidence in heterogeneous forms: as a text passage, as a rendered image of the same passage, or as both together. However, it remains unclear how consistently these surface forms are processed, especially when the evidence conflicts with the model's parametric knowledge. We study modality robustness under knowledge conflict across 13 MLLMs and two datasets, and find them far from robust. (1) Contrary to common belief, models favor a context that contradicts parametric knowledge more readily in image form than in text form; (2) when a contradicting text and image are presented together, the preferred modality is essentially arbitrary, varying with input order, model, and dataset. We further demonstrate that this instability has practical consequences: it degrades performance in multimodal RAG and can be exploited by adversarial attacks. To alleviate this brittleness, we examine several simple techniques—prompting, steering, supervised fine-tuning (SFT), and direct preference optimization; the majority prove ineffective, whereas SFT achieves moderate success. We therefore call for greater awareness of this inconsistency and argue that it is fundamental, demanding attention at multiple training stages.

Large Language Model (LLM) agents increasingly rely on external skills, yet standard evaluations obscure whether retrieving these skills actually helps. Aggregate metrics often compare retrieved versus non-retrieved tasks, introducing severe selection bias and failing to isolate the true effect of skill use. To measure this actual-use capability-which we formalize as Skill Following (SF)-we introduce the Retrieval-Invoked Actual-Use Effect (RAE). RAE computes the same-task outcome difference between matched skill-enabled and skill-disabled executions, conditioned exclusively on tasks where the agent actively retrieved a skill. Evaluating 17 LLMs across coding and mathematical domains, we uncover a stark evaluation paradox: models frequently show positive aggregate retrieval lift but negative RAE. On MBPP+, multiple models that appear to benefit system-wide actually harm their own performance on the exact tasks where retrieval occurred. These findings demonstrate that aggregate averages can create a misleading illusion of tool-use proficiency, whereas RAE directly measures whether the retrieval-to-answer pipeline genuinely rescues more outcomes than it harms.

Large language models (LLMs) have recently demonstrated improved machine translation performance over strong supervised baselines. This raises questions as to what mechanisms underlie how LLMs perform machine translation between languages. Motivated by recent interpretability findings–namely, that LLMs use massively multilingual latent feature representations to perform language modeling–we propose the interlingua hypothesis. The hypothesis holds that language models translate by reading a source sentence into a latent feature space, and generate a target sentence by reading from the latent feature space. We show three lines of evidence in support of this hypothesis: (1) variance in BLEU across language pairs is largely predictable from language-specific competences with no language pair-specific interaction terms; (2) many model components are causally influential in both monolingual tasks and translation tasks; and (3) fine-tuning on monolingual data recovers a large proportion of translation improvements relative to fine-tuning on aligned documents. Together, these provide convergent evidence in support of the interlingua hypothesis, and suggest new ways of understanding and improving how LLMs can be leveraged to perform translation tasks.