LLMs

Researchers analyzed 367,174 MCP tools and found that functionally similar tools often have divergent parameter schemas despite sharing names and descriptions. By conducting controlled interventions across five LLMs, the study isolates how specific metadata factors like schema structure influence tool selection decisions. The findings confirm that parameter schema differences significantly affect which tool an LLM chooses before any execution occurs.

• Functionally identical MCP tools can have vastly different parameter schemas.
• LLMs select tools based on metadata, making schema structure a critical factor.
• Controlled interventions prove schema variance directly alters LLM tool selection.
• Standardizing metadata descriptions alone may not ensure consistent tool routing.
• Schema normalization is essential for reliable MCP agent behavior.

Research demonstrates that autoregressive generative recommendation systems are susceptible to conditional targeted poisoning attacks. By injecting fewer than 1% of synthetic interactions as malicious data, an adversary can significantly increase the likelihood of promoting specific targets when triggered. The study highlights a critical resilience gap in these flexible models compared to classical ranking systems.

• Generative recommenders are vulnerable to targeted poisoning via injected synthetic interactions.
• Attacks succeed with less than 1% malicious data injection under black-box API conditions.
• Trigger-based prompting can significantly skew recommendation outcomes toward malicious targets.
• Resilience of generative models to data poisoning remains largely unexplored and risky.

Next-Concept Modeling (NCM) is a two-stage framework that uses a multiscale residual VQ-VAE to encode text into a hierarchy of discrete codes. A transformer then generates these codes level by level, predicting all codes within a specific level in parallel. This approach aims to combine the speed of parallel sampling with high-quality generation, addressing the degradation issues seen in diffusion language models at low sampling steps.

• NCM uses a multiscale VQ-VAE to create a hierarchy of discrete text codes for coarse-to-fine generation.
• A transformer predicts all codes within a level in parallel, enabling faster inference than autoregressive models.
• NCM achieves better perplexity than top diffusion baselines on LM1B while running 21x faster.
• The method mitigates quality degradation at low sampling steps common in diffusion language models.

This research paper argues that reported reasoning improvements in hybrid models like Qwen3, trained via reinforcement learning in NO-THINK mode, are largely due to thinking leakage. Instead of learning genuinely enhanced reasoning behaviors, the models partially reactivate the original THINK mode's capabilities. Evidence shows that as accuracy improves, responses become longer, more reflective, and increasingly resemble the model's internal thinking patterns.

• RL training in NO-THINK mode may not create new reasoning skills but rather reactivate latent THINK mode abilities.
• Improved accuracy metrics correlate with response lengths and styles that mimic the original model's thinking process.
• Practitioners should verify if performance gains stem from genuine behavioral shifts or leakage from the THINK mode.
• Internal activation shifts toward THINK-mode patterns suggest the mechanism is reactivation, not novel learning.

Researchers introduce EdgeLLM-Nano, a measurement study using an 8.39M parameter decoder-only transformer to evaluate KV-cache compression on edge devices. The study highlights that existing evaluation methods often fail to include the KV-cache under test, leading to inaccurate assessments. Results deliver a negative verdict for polar-guided key-cache quantizers, indicating significant challenges for current compression techniques on nano-scale models.

• EdgeLLM-Nano uses a 33.6 MB model to enable end-to-end examination of backend, seed, and power samples.
• Existing KV-cache evaluation methods are flawed because they often do not include the KV-cache under test.
• Polar-guided key-cache quantizers received a negative verdict, suggesting poor performance for this approach.
• Memory, latency, and power constraints on edge devices require specialized compression strategies beyond HPC standards.