The editorial team selects four papers from the arXiv submissions of the past 24 to 48 hours that demonstrate how differently capability, trustworthiness, and efficiency of today’s AI systems currently diverge – from a new open large model to tipping safety panels and freely invented user profiles to more tightly controlled thought traces. The curation was based on substance: each paper offers a traceable method and a concrete numerical result in the abstract, not just a topic headline. The selection deliberately covers four different subfields – model release, safety assessment, personalization, and inference efficiency.
K-EXAONE 2.0: LG AI Research scales to 750 billion parameters
A large author team from LG AI Research presents K-EXAONE 2.0, an open multilingual foundation model that was not retrained from scratch but expanded from its predecessor K-EXAONE through “upcycling”: a mixture-of-experts model (MoE, a model that activates only part of its specialized subnetworks per request) with 750 billion total and about 37 billion active parameters per token – more than three times the capacity of its predecessor. The model supports context lengths of up to 256,000 tokens and grows from six to ten supported languages. The training pipeline combines continued pre-training, difficulty-focused mid-training, and post-training to strengthen reasoning, agentic coding, multilingual capability, and safety grounded in Korean societal contexts. Across nine practically chosen evaluation categories, K-EXAONE 2.0 reportedly improves over its predecessor and remains competitive with open-weight models, with the largest gains in agentic coding and long-context understanding. The model is released under the open Apache-2.0 license. This matters because it is another very large open MoE model with concrete parameter and context numbers from an Asian lab – a pattern seen before with Thinking Machines’ model Inkling, which has 975 billion total and 41 billion active parameters, except that K-EXAONE 2.0 additionally ties safety explicitly to a specific cultural context.
Safety panels: one false peer tips the entire majority vote
Yibo Hu and Jiaming Qu examine in Social Pressure Breaks Majority Voting in LLM Safety Panels what happens when several AI models are meant to jointly flag unsafe content as a panel but see the same misleading cues before voting. In a two-round experiment, each model first judges a case alone, then again after six simulated peers either assert a wrong label or abstain; the final judgments are combined by majority vote. Across six open-weight language models and six datasets, the authors report that the average false-alarm rate of individual reviewers rises from 56.5 percent (with silent peers) to 87.5 percent once a wrong label is asserted, and the panel’s majority vote drives the false-alarm rate up to 100 percent. Without an asserted label, the same panel outperforms its average member. The effect is strongly asymmetric: reviewers follow a push toward “unsafe” far more often (about 75 percent) than toward “safe” (about 17 percent), so the false-alarm rate rises sharply while the rate of missed genuine risks barely changes. This matters because companies increasingly deploy model panels as an extra safety layer – a follow-up to the previously discussed audit of agent-safety benchmarks, which already showed how unreliable common safety measurements can be, though there through flawed test construction rather than, as here, through shared social cues among the reviewers themselves.
Fake profiles: AI invents user traits – and misjudges its own reliability
Yushi Sun, Yanjie Zhang, and Rui Sheng investigate in The Personalization Mirage: How LLMs Fabricate User Profiles, and Why Self-Monitoring Misleads the phenomenon of “over-inference”: language models with persistent memory fabricate user attributes beyond what the available evidence supports. For the new MirageBench benchmark, the authors use 150 personas spanning stereotypical, counter-stereotypical, and neutral profiles, six personalization tasks, and a four-way faithfulness taxonomy validated against a blind human annotator on 400 claims (Cohen’s kappa 0.863 in the four-class version); in total, an independent judge model evaluates 143,616 individual claims from twelve models across seven model families. Over-inference proves pervasive: every one of the twelve models fabricates 35 to 49 percent of its claims (cross-model mean 41.6 percent), with no model escaping the pattern. Most strikingly, the authors report a “self-monitoring inversion”: at the model level, a model’s self-assessed over-inference rate correlates negatively with the rate measured by the judge model (rank correlation ρ = -0.60) – the models that report the least over-inference tend to be caught fabricating the most, even though within a single model, self-auditing still ranks that model’s own claims moderately well (AUROC 0.58 to 0.83). This matters because personalized assistants with long-term memory are increasingly deployed in everyday use, and the study shows that trusting a model’s own report about its reliability is precisely the signal that misleads here.
ReCo: fewer tokens, smaller cache through reward-coordinated compression
A team led by Qiyuan Zhu and Dezhi Li presents Fewer Tokens, Smaller Cache: Reward-Coordinated Efficient Reasoning, the ReCo framework, which combines two approaches to speeding up reasoning models that have so far been handled separately: KV-cache compression (a technique that shrinks the intermediate state the model stores during generation) and control of the raw token count. The authors observe that individual reasoning steps differ in how sensitive they are to context loss, and that an overly aggressively shrunk cache can paradoxically push the model to generate more text rather than less. A lightweight process-reward estimator therefore scores each completed reasoning step and jointly steers three mechanisms: adaptive KV-cache compression that intensifies at high-reward steps, reward-based penalties on redundant reflection tokens, and confidence-driven early stopping. Tested across three reasoning models and six benchmarks, ReCo reportedly cuts the number of generated tokens by 37 to 65 percent and speeds up responses by a factor of 2.08 to 2.35 relative to full chain-of-thought, while substantially preserving accuracy. This matters because it marks a direct advance over the previously covered method BLADE, which saves tokens purely through early exit: ReCo additionally couples that exit logic with the memory side of the problem.
A closing note on context: all four papers are still unreviewed preprints, and their figures come from the respective author teams’ own experiments. The K-EXAONE evaluation relies on nine self-chosen evaluation categories, the panel study covers six open-weight models and six datasets, MirageBench covers twelve models with a fairly wide confidence interval around its central correlation, and ReCo has so far only been tested on three models and six benchmarks. Whether these patterns hold up across more models, tasks, and independent replications remains to be seen.
