Persistent web agents improve over repeated tasks only if they know user preferences that history alone cannot reveal. This paper formalizes Active Preference Acquisition (APA), where an agent decides which missing attributes to ask about by estimating their Prospective Personalization Value: the expected cumulative utility gain across future tasks minus asking and privacy costs. The core insight is that many preferences are worth acquiring only when reused across many later tasks, since no single booking or purchase may justify the cost of asking.

Monolingual language models predict native reading times well but miss how a reader's first language shapes second-language processing. This paper investigates whether bilingual GPT-2 models sequentially trained on an L1 and then English can capture human cross-linguistic influence during L2 reading. The authors evaluate 55 models---monolingual and bilingual with varying steps of L1 exposure---against eye-tracking data from 12 L1 backgrounds, showing that L1-matched bilingual surprisal improves prediction of late-stage reading measures and peaks for intermediate-proficiency learners. The work is significant because it provides a computational bridge between multilingual NLP and psycholinguistic theories of non-native sentence processing.

This survey tackles long-document visual question answering, where answers require locating and integrating evidence scattered across many pages and modalities including charts, tables, and fine-grained layout. Rather than grouping papers by model architecture, the authors decompose the task into four functional components: document parsing, evidence organization and retrieval, generator organization and learning, and evaluation. The core argument is that task bottlenecks—visual fidelity, evidence dispersion, and budget limits—should drive taxonomy design, not just technical novelty. This reframing matters because the field has grown explosively since 2024 and prior surveys have organized literature by approach rather than by the functional requirements that long documents impose.

ReflectPO tackles autonomous error recovery in long-horizon LLM agents, where early execution mistakes compound across sequential tool-use interactions and existing RL methods optimize task outcomes without teaching the policy when and how to self-correct. The core idea is to generate recovery trajectories from rollback snapshots, guided by reflection hints derived from branch differences across rollout trees and repeated loops within trajectories, and then optimize the policy with tree-structured credit assignment that captures both local decision effects and recovery gains. This matters because preliminary evidence shows that errors are nearly ubiquitous in natural trajectories yet agents autonomously recover in fewer than 15% of cases, leaving substantial headroom for training-based correction.

Knowledge graph-to-text generation converts structured triples into fluent natural language, yet whether modern prompted LLMs can match task-specific fine-tuned models remains unclear. This paper benchmarks five recent model families—dense, mixture-of-experts, and encoder-decoder architectures—against a fine-tuned T5-large baseline on WebNLG and DART, using eight reference-based metrics together with a graph-grounded LLM judge validated by human annotation. The core finding is that prompted LLMs have nearly closed the gap on traditional metrics while exhibiting stronger faithfulness and fewer omissions on larger graphs, underscoring that surface similarity scores alone are no longer sufficient to assess generation quality.

Crisis warning generation is currently English-centric, hampering emergency response in linguistically diverse Europe. This paper introduces CrisiLangs, a multilingual dataset extending 1,000 expert-reviewed English event-message pairs into Italian, Bulgarian, German, and Greek via professional native speakers. The authors benchmark the resource by jointly fine-tuning Llama-3.1-8B-Instruct across all five languages, showing gains over zero-shot baselines but also revealing that generation quality and the utility of in-prompt guidelines vary substantially by language.

Speculative decoding draft models face severe KV cache bottlenecks during long-output generation, but standard prefill-based compression methods lose relevance as decoding progresses. This paper identifies that "decode staleness" phenomenon and proposes RIFT, a family of selection-based KV compression strategies that balance important prefill tokens with a recent-token sliding window, plus adaptive decode-time refresh and a complementary compression-aware distillation training phase. The core claim is that these methods allow small external drafters to retain over 96% of their full-cache acceptance length under tight memory budgets.

Fine-grained image-text retrieval requires aligning local semantics such as objects, attributes, and relations, yet fixed patches and tokens often produce fragmented matching units. This paper proposes HFCMA, a framework that dynamically aggregates low-level patches and tokens into adaptive visual regions and textual phrases, and then aligns these sets with entropy-regularized optimal transport under a frozen CLIP dual-encoder architecture. By making semantic unit construction an explicit learnable step and enforcing globally consistent matching via a transport plan, HFCMA aims to improve discrimination of subtle differences without sacrificing retrieval scalability.

Guardrails for LLMs typically fail when the same safety case appears across different interfaces—prompts, responses, agent traces, or policy-conditioned inputs. UniGuard reframes guardrail training as multi-view representation learning, where prompts, conversations, and agent trajectories are treated as semantically aligned views of an underlying safety case. A compact 0.6B model is trained with supervised classification, contrastive alignment, and asymmetric confidence calibration, showing large F1 and calibration gains on unseen-interface benchmarks while remaining competitive with proprietary models.

Language models fail to learn shortest-path search on unseen symbolic graphs as graphs scale beyond about 50 nodes, a wall this paper diagnoses and breaks through. The authors show that next-token prediction provides too little and too skewed supervision, while standard tokenization introduces routing and permutation artifacts. By combining multi-token prediction—which densely supervises all sub-paths and implicitly mixes hop lengths—with structural annotation (STAN) and a node list (+NN), they stably train small transformers on graphs up to 125 nodes with far lower variance. The work makes a strong case that objective design and input formatting matter more than model size for latent multi-hop reasoning.

The paper studies how much embedding work is wasted when documents in a RAG knowledge base are revised, comparing fixed-size, recursive paragraph-first, and content-defined chunking on 120 adjacent Wikipedia revision pairs in Chinese and English. It finds that fixed-size windows force the most re-embedding (57–68\% of the embedding-input workload), while content-aware methods cut this by double-digit percentage points. Surprisingly, the extra recomputation buys almost no retrieval recall advantage under a fixed 2048-token budget. The authors argue that update cost, measured under byte-identical reuse, should be reported alongside static retrieval metrics when choosing a chunker for an evolving knowledge base.

Hierarchical language models split semantic planning from token generation, but prior work leaves the latent space and decoder mismatched. Cola2 co-trains a dynamic compression VAE and a masked-diffusion decoder so that token-prediction gradients shape the encoder while a separate DiT fits the latent prior via stop-gradient. The result is a compact continuous representation that stays aligned with the actual text during block-wise inference, and scaling experiments up to about 210B tokens show competitive, unsaturated curves. A 346B-source-token checkpoint reaches a 60.10 eight-task average, though the model remains far from converged.

LLM agents with persistent external memory often leave derived artifacts—such as summaries—intact after a user deletes a source record, creating a gap between storage-level deletion and behavioral privacy. AfterDeleteBench introduces a paired behavioral audit that compares an exposed-then-deleted agent against a never-exposed twin to attribute post-deletion secret recovery to the original exposure. The paper shows that exact reconstruction can persist even when no verbatim copy of the secret survives, and that the response interface itself is part of the measurement system, making the benchmark a useful narrow probe into a broader agent-memory privacy problem.

Existing cultural bias benchmarks rely on costly manual annotation by native speakers, which limits multilingual scalability and introduces subjectivity. WikiOriginQA tackles this by using Wikipedia article creation dates across 342 language editions as an objective proxy for knowledge origin, then automatically generating short-answer QA pairs with LLMs. The resulting dataset supports the claim that LLMs are skewed toward Anglophone knowledge and that propagation into English Wikipedia is a key predictor of model performance.

The paper challenges the widespread assumption that reinforcement learning (RL) invents the reflection marker "wait" in reasoning models. Using historical-linguistics-inspired "grammaticalization" tracking, the authors trace the word through 120 open-model checkpoints and find it almost always arrives from training data—mid-training, fine-tuning on reasoning traces, or reasoning-rich pretraining—while zero-RL alone rarely coins it.

Extraction pipelines often ask language models to return JSON matching a schema that sets a minimum list length. This paper shows that when real documents contain fewer items than that minimum, open-weight models pad their lists---mostly with real text from the document that evades simple source checks. The authors preregister a design with 558 documents, eight GPU models, and paired contrasts, finding that an unenforced minimum raises over-counting from 34.5% to 55.4% and that common fixes fail or trade recall for precision. The work matters because schema-driven extraction is ubiquitous, and the padding is hard to catch without ground-truth answer lists.

On-policy self-distillation improves reasoning by training a model on its own trajectories with dense token-level guidance from a reference-conditioned teacher. The authors identify a critical flaw: because any single reference is just one valid realization of reasoning, conditioning the teacher on the full reference makes the target distribution overly sensitive to that specific solution's wording, ordering, and path. They propose Monte Carlo Privileged Distillation (MCPD), which averages the teacher's next-token distributions over multiple randomly masked partial views of the same reference to produce a more stable marginal target. Across math and science benchmarks, MCPD consistently outperforms standard full-reference distillation.

Nepali–English code-switched ASR suffers from script confusion and switch-point fragility, yet human annotated conversational audio is costly. This paper curates a 15.45-hour spontaneous podcast benchmark and adapts a 1.2B-parameter multilingual encoder–decoder model using only 0.95 hours of synthetic dual-engine TTS data. The core finding is that supervised fine-tuning drives almost all cross-domain transfer, while subsequent GRPO and script-aware rewards add only marginal, hard-to-disentangle gains. The released data and the cautionary ablation make the work practically useful.

Coding-agent benchmarks increasingly grade agents against upstream test suites that are paired with natural-language specifications, yet those tests often require unstated names, module paths, or literals the specification never gives. This paper formalizes that mismatch as the specification–oracle gap and implements a fast, model-free static audit for five languages, then shows that on SWE-Milestone roughly a quarter of graded milestones contain hard hidden dependencies that can block entire build units. The finding matters because it reveals that many benchmarks inadvertently measure whether agents reproduce upstream implementation details rather than whether they satisfy the stated requirements.

The paper tests whether open LLMs that a hospital could self-host are ready for clinical use, evaluating four models (8–27B parameters) on the Spanish MIR residency exam under its real scoring and ranking rules. The authors go far beyond headline accuracy to probe calibration, the value of abstention under a wrong-answer penalty, the soundness of clinical reasoning, and whether vision models actually use the images they receive. Their headline message is that a model can rank high enough to choose almost any specialty while remaining overconfident, correct for the wrong reasons, and blind to the clinical plates it is shown.