TopicMerge tackles the scalability bottleneck of LLM-based topic modeling by combining efficient clustering with targeted LLM reasoning. Its core idea is to first over-segment a corpus into fine-grained semantic clusters, then use a lightweight LLM to hierarchically merge them with contextual neighborhood guidance. The authors claim this slashes LLM prompts by over 92% while matching or exceeding state-of-the-art clustering alignment on standard benchmarks.

Most benchmarks test whether LLMs can produce or understand mixed-language text, but not whether they decide when, where, and why to switch for narrative effect. CSWBench fills this gap with a creative-writing benchmark that pairs focused and extended narrative generation with a theory-informed evaluation framework assessing event-level licensing, contextual fit, and functional gain alongside whole-text multilingual orchestration. It matters because code-switching in stories should be a deliberate narrative resource, not just surface multilingual decoration.

Standard Character Error Rate charges every OCR edit equally, whether it is a harmless typo or a silent factual error like changing a name. This paper introduces semCER, which keeps CER's grapheme alignment and denominator but weights each edit by a severity score from an LLM judge, eliminating the need for word segmentation. Validated against human readers in English, Vietnamese, and Chinese, semCER better ranks damaging errors and breaks ties where CER sees identical edit counts, offering a more nuanced tool for OCR and vision-language model evaluation.

Corporate loan applications must comply with dense internal lending policies, but manual review is knowledge-intensive and prone to missed exceptions. This internship report develops a two-step LLM pipeline that first extracts machine-readable rules from policy documents and then evaluates real loan applications against them. Tested on eleven non-anonymized corporate loans at Erste Bank, the system detects most known policy exceptions but shows significant limitations in rigorous validation and scalability.

The paper tackles the scarcity of multilingual safety and reasoning benchmarks for LLMs by proposing a reasoning-guided translation framework. It separates translation strategy selection via a reasoning model from large-scale generation by a cheaper translator, then validates, repairs, and independently screens outputs in a task-specific manner. Using seven English sources across 11 target languages, the authors construct a source-aligned collection and expose language-dependent vulnerabilities such as inconsistent refusals and substantial reasoning degradation.

Autoresearch LLM agents plateau when their growing session history makes them conservative. This paper shows that periodically discarding everything except the best candidate—forgetting—resets exploration and unlocks further gains. The finding suggests that how an agent manages memory is as important as how much compute it receives.

SwarmRSI tackles recursive self-improvement in scientific research by treating hypotheses, programs, and experimental evidence as inheritable states that evolve across a swarm of LLM agents. Its core idea is a shared research graph enabling feedback-guided mutation, recombination of partial solutions across agents, and adaptive allocation between hypothesis and program refinement. The framework reports state-of-the-art scores on OpenProblems and NatureBench and provides ablation and trajectory evidence that cross-agent reuse drives progress.

Video large language models often hallucinate answers that are not supported by the input footage. Contrastive decoding can reduce this by comparing the original video against a weakened "negative" view, but existing negatives are either predefined or optimized for query reconstruction rather than the model's actual visual reliance during answer generation. ARP-CD introduces a training-free fix: it treats the attention mass on visual tokens at the answer-prediction position as a differentiable optimization target, then uses bounded gradient ascent to perturb the input visual hidden states and create a negative branch for contrastive decoding. The authors evaluate the idea across four Video LLMs and four benchmarks, supplementing headline results with ablations and a mechanistic analysis of first-token logit margins.

PerSo introduces a benchmark for evaluating LLM social intelligence through real-world persuasion dialogues from Reddit's ChangeMyView. It tests cognitive intelligence by forecasting a persuadee's beliefs, emotions, and intents, and behavioral intelligence by selecting the more persuasive response. Unlike prior work that often relies on synthetic scenarios or isolates these two aspects, PerSo jointly evaluates both on naturally occurring interactions, providing a more ecologically valid testbed.

This paper identifies a gap between standard visual grounding—which assumes enough dialogue context to pinpoint a single object—and real-world interactions where multiple referents remain plausible. It proposes Candidate Set Visual Grounding (CSVG) and introduces the CandVG benchmark to evaluate how the set of plausible referents evolves across dialogue turns. Evaluating 18 vision-language models shows that stronger single-object grounding does not translate to better set estimation, with models often failing to generate full candidate sets or committing prematurely to one candidate.

Multimodal code intelligence demands programs that are correct, well-structured, and visually faithful, yet existing training signals are either too coarse or confined to single tasks. This paper introduces PrismaForge, which synthesizes per-instruction evaluation rubrics that separately assess code-level and visual-level criteria, then uses them to filter training data and reward reinforcement learning. The resulting dataset, PrismaCoder-2M, and the trained 9B and 35B models advance the open state of the art and compete with proprietary systems on charts, webpages, and scientific graphics.

This paper investigates how Retrieval-Augmented Generation affects question answering across five South Asian languages. The authors construct a Wikipedia-based multilingual dataset, generate questions with Qwen 3 32B, and compare non-contexted, contexted, and RAG-based answering using four LLMs. They find that RAG improves QA unevenly—substantially aiding non-popular cultural questions in Bangla and Tamil while underperforming direct contexted answering in Hindi and Nepali—suggesting that retrieval quality and language-specific data alignment matter more than the choice of RAG framework.

Current social agent benchmarks reward fluent conversation but ignore whether agents actually reach valid plans that respect hidden constraints. This paper introduces HICO-Bench, a benchmark where two agents with private information must coordinate verifiable outcomes, and PaCPO, a reinforcement learning method that redistributes terminal rewards across turns by checking tentative plans against paired worlds. The work matters because it reveals that high LLM-judged conversation ratings frequently mask infeasible or unsupported agreements, and provides a concrete credit-assignment mechanism to train agents on verified social outcomes rather than just dialogue style.

LLM entity extraction from full-length documents is widely deployed but its failures are poorly diagnosed. This paper runs a preregistered audit of two low-cost LLMs and a flagship control on 315 real-world cyber-threat-intelligence reports, finding that most false positives are background mentions rather than hallucinations, and that confidence signals can rank outputs for human review but cannot verify truth against relation-level counterfactuals. The results set a concrete deployment boundary for any high-stakes pipeline that triages extraction by confidence.

This paper studies how to convert U.S. state legislative PDFs into exact, schema-constrained XML. The authors test the hypothesis that exact reproduction relies more on harness engineering (tools, orchestration, deterministic constraints around a model) than on model scale. They cross three Claude models with three harness tiers (zero-shot, single-agent with tools, and decomposed multi-agent pipeline) on 153 gold bills. The headline finding is striking: the weakest model in the full pipeline outperforms the strongest model zero-shot ($0.948$ vs $0.806$), and the model spread collapses from $0.68$–$0.81$ to $0.95$ under full scaffolding.

On-policy self-distillation gives reasoning models dense token-level supervision via a teacher that sees reference solutions, but this signal can drift toward style and ignores whether the student can finish correctly on its own. VASPO repurposes the same privileged reference inside PPO to train both a prefix-value critic and a bounded self-teacher that reweights selected positive advantages. The paper reports average gains of roughly 2.6 percentage points over PPO on math benchmarks and leading accuracy on Knights and Knaves.

Most synthetic pretraining pipelines generate samples by conditioning on an entire source passage or paragraph and asking a model to rewrite or paraphrase it. The paper identifies a systematic failure mode of this approach: fine-grained facts—such as secondary entities, list and table entries, and precise temporal attributes—are frequently omitted because the generator implicitly selects which facts to preserve. Atomic-Fact-Conditioned Synthesis (AFCS) fixes this by first extracting source-supported atomic facts, then conditioning each synthetic training sample on an explicit retained fact and its local context. Continued-pretraining experiments on 7B and 32B models show consistent gains on factual QA benchmarks, especially for middle- and long-tail knowledge, and granularity ablations isolate synthesis granularity as a causal lever for factual recall beyond surface-form diversity.

Probes on hidden activations are a promising tool for detecting deception and harmful intent, but they assume models cannot volitionally alter their own representations. Baldelli et al. ask whether open-weight LLMs can control their activations through in-context neurofeedback, devising a cross-model writer-subject protocol to disentangle genuine self-access from in-context pattern matching. Across 42 semantically opaque readouts and six safety monitors, they find that free-generation effects are largely explained by sentence-selection competence, while fixed-sentence control is tiny and monitor evasion is essentially absent. The results rigorously bound the current threat but flag a capability that may grow with model scale.

Agent skills let frozen models adapt through editable external instructions, but optimizing them faces a tension: batching many changes obscures which edit mattered, while judging each edit only by its immediate score risks discarding useful stepping stones. AtomSearch proposes contrastive atomic editing—one primitive rule edit tied to execution evidence—and Skill-State Tree Search (STS), which tests short edit compositions and keeps a composed state only if it beats its primitive child. The hope is to find compact, inspectable skills without greedy myopia.

Standard ReAct agents carry every failed attempt and off-path observation in one linear context, which eventually buries the evidence they need. TreeFlex solves this by saving intermediate states as nodes in a search tree and letting the same language model choose where to resume and what to do next, operating in two roles—Reflection for global navigation and Execution for local work—trained end-to-end with a tree-aware adaptation of GRPO. The paper shows that this design improves accuracy on deep research and software engineering benchmarks, especially on harder tasks, though it comes with higher inference costs and small-scale training.