Also Worth Noting - 2026-06-09

Also Worth Noting

02 [Theory] On the Geometry of On-Policy Distillation On-policy distillation does not behave like a cheaper version of RLVR. Parameter-space diagnostics consistently place OPD in a relaxed off-principal regime: its updates touch fewer weights than SFT and avoid principal directions more strongly, yet remain less tightly constrained than RLVR. The geometry sits between the two methods, not on top of either. Teams treating OPD as a drop-in RLVR substitute are optimizing a structurally different objective than they think, which matters when choosing training recipes for reasoning models. link

03 [Agent] Send a SCOUT First: Pre-hoc Reasoning for Adaptive Detector Allocation in Prompt-Injection Defense Fixed single-detector pipelines for prompt injection commit every request to one detector's blind spots, and no single detector is reliable across all attack slices. SCOUT reframes defense as per-request detector allocation: it predicts each detector's reliability and latency for the incoming sample before deciding which detectors to run, and escalates to an LLM judge only when uncertainty warrants it. The result is systematic blind-spot reduction without running every detector on every query. Anyone building production LLM security pipelines should treat this routing-first architecture as the new baseline. link

04 [Inference] FlashMemory-DeepSeek-V4: Lightning Index Ultra-Long Context via Lookahead Sparse Attention Rather than attending to all past tokens, Lookahead Sparse Attention proactively predicts which KV chunks will be needed next and keeps only those in GPU memory. A Neural Memory Indexer, trained with a backbone-free decoupled strategy on top of DeepSeek-V4, handles the prediction without retraining the base model. The approach directly attacks the KV cache memory bottleneck that makes ultra-long context serving expensive at scale. For inference infrastructure teams, this is a prefetching-style solution to a problem that brute-force context windowing cannot solve. link

05 [Training] Reasoning Arena: Trace Tournaments When Verifiable Rewards Fall Short When every rollout for a given prompt receives identical rewards, group-relative advantage estimation produces zero gradient signal and RLVR training stalls entirely, even when traces differ substantially in reasoning quality. Reasoning Arena detects these degenerate reward groups and routes them to a tournament-style judge that generates relative rankings across traces, recovering a usable learning signal without modifying the reward model. The fix targets a failure mode that becomes more common as models improve and correct answers get easier to hit. Teams running RLVR at scale should expect this problem to grow, not shrink. link

06 [Open-source] i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models Unlike Flux or SD3, which release weights without training data or full hyperparameters, i1 publishes weights, data, and code together, making it the first competitive text-to-image baseline where ablations are actually reproducible. The project runs systematic ablations on modeling and data design choices to isolate what drives recent diffusion progress, rather than attributing gains to opaque combinations of scale and undisclosed curation. That transparency is the contribution as much as the model itself. For researchers trying to understand why diffusion quality improved over the past two years, i1 provides the controlled foundation that closed models never offered. link