Also Worth Noting - 2026-06-14

Also Worth Noting

02 [Agent] Control-Plane Placement Shapes Forgetting: An Architectural Study of Agent Memory Across Thirteen System Configurations Where the LLM sits in an agent memory pipeline determines which forgetting failures the system can recover from -- and most frameworks leave that placement decision implicit. Across thirteen configurations tested on a 385-case adversarial surface, systems where the LLM sits before the control plane (write operations) recover substantially better from supersede and purge failures than those placed after. Deterministic primitives handle lexical and temporal categories but collapse on canonicalization, hitting 5% on identifier-obfuscation and 0% on cross-lingual cases. Teams building production memory layers should treat LLM placement relative to the control plane as a first-class architectural choice, not a default. link

03 [Eval] Formalize Once, Edit the Rest: Efficient Lean-Based Answer Selection for Math Reasoning Lean-based answer selection at test-time scaling has a K-fold formalization cost problem: existing approaches generate a separate formal statement for every candidate answer, making verification expensive at scale. Formalizing only one candidate and propagating the Lean proof to filter the remaining K-1 candidates cuts that overhead by roughly a factor of K. The approach preserves machine-checkable rigor without multiplying autoformalization calls. For teams running test-time compute scaling on math reasoning, this makes formal verification practical rather than prohibitively expensive. link

04 [Eval] Who Flips? Self- and Cross-Model Counterarguments Reveal Answer Instability in LLMs Models that score identically on accuracy benchmarks can differ by more than 20 percentage points in answer stability when a plausible counterargument challenges a correct response. The protocol isolates argumentative content from social pressure and varies argument length and source, exposing a reliability gap that standard benchmarks cannot see. A model that flips correct answers under coherent opposition is not deployment-ready in adversarial or debate-style settings, regardless of its accuracy score. Stability under challenge should sit alongside accuracy as a standard evaluation axis. link

05 [Training] PreLort: Prefix-Nested LoRA for Federated Fine-Tuning under Rank Heterogeneity Heterogeneous hardware in federated fine-tuning means clients operate at different LoRA ranks, and current aggregation methods cannot control how information distributes across rank dimensions -- leaving shared low-rank representations underused. PreLort fixes this by nesting lower-rank client adapters as prefixes of higher-rank ones, so aggregation across mismatched hardware is direct and the information-leakage problem that plagues existing heterogeneous federated LoRA approaches is eliminated by construction. Teams running federated LLM fine-tuning across mixed-capability edge hardware now have an aggregation scheme that respects rank structure rather than discarding it. link

06 [Open-source] Open-SWE-Traces: Advancing Dual-Mode Multilingual Distillation for Software Engineering Agents The data bottleneck for multilingual coding agents just got a direct answer: 207,489 agentic trajectories sourced from 20,000 real-world pull requests across nine languages including Rust, Go, TypeScript, and C++. A hybrid synthesis pairs Minimax-M2.5 for explicit chain-of-thought traces with Qwen3.5-122B for high-quality non-thinking traces, covering both reasoning modes in the same dataset. All trajectories are filtered for permissive licenses. This directly addresses why multilingual coding agent research has stayed concentrated at well-resourced labs -- the training data simply was not publicly available at this scale before. link