Also Worth Noting - 2026-06-25

Also Worth Noting

02 [Training] Improved Large Language Diffusion Models Masked diffusion at 8B parameters and 12T pre-training tokens can match autoregressive models on several benchmarks, which puts real pressure on the assumption that AR is the only viable path at scale. iLLaDA keeps the masked diffusion objective through both pre-training and supervised fine-tuning, scaling SFT on a 25B-token instruction corpus over 12 epochs with variable-length generation for efficiency. Confidence-based scoring replaces the standard multiple-choice evaluation convention. Teams designing next training runs should treat AR factorization as a choice, not a constraint. link

03 [Inference] RoPE-Aware Bit Allocation for KV-Cache Quantization Treating a cached key as a flat vector is the wrong abstraction under RoPE: a key's contribution to an attention logit decomposes into a position-dependent sum over two-dimensional frequency blocks, making some blocks far more sensitive to quantization error than others. Block-GTQ assigns bits per RoPE frequency block using a label-free energy score, allocating more bits where quantization damage is highest. The result is lower KV-cache memory without the accuracy loss that has made aggressive uniform-bit quantization risky in production long-context deployments. link

04 [Agent] Constraint Tax in Open-Weight LLMs: An Empirical Study of Tool Calling Suppression Under Structured Output Constraints Multiple open-weight models stop calling tools entirely when JSON Schema output constraints are active at the same time, even while maintaining high schema compliance. This "tool suppression" is reproducible across model families and deployment settings, meaning it is a structural interaction failure, not a one-off model quirk. Teams shipping production agent systems with both structured output and tool calling enabled should audit whether tool invocation rates have silently collapsed rather than assuming schema compliance confirms correct behavior. link

05 [Eval] When Lower Privileges Suffice: Investigating Over-Privileged Tool Selection in LLM Agents LLM agents consistently select higher-privilege tools than a task requires, even when a lower-privilege alternative would suffice. ToolPrivBench measures both initial tool choice and privilege escalation behavior across agents operating autonomously, exposing a concrete security surface that standard tool-selection benchmarks ignore entirely because they optimize for capability, not privilege sensitivity. Teams deploying autonomous agents in environments with tiered tool permissions should treat over-privileged selection as a default behavior to constrain, not an edge case to monitor. link

06 [Theory] Diagnosing Task Insensitivity in Language Agents LLM agents fail out-of-distribution tasks not from general capability gaps but from pattern-matching to training-time task templates. When a task description is semantically corrupted or swapped for a similar but distinct task, models continue executing actions aligned with the original template rather than adapting. This means evals that only test in-distribution task variants are systematically hiding a known generalization failure. Any agent benchmark that does not include semantically perturbed variants of familiar tasks is measuring memorization coverage, not generalization. link