Also Worth Noting - 2026-06-18

Also Worth Noting

02 [Training] STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability GRPO entropy collapse traces to a token-level credit assignment mismatch, not a reward design problem. High-surprisal tokens dominate gradient updates and destabilize policy entropy because per-token entropy variation decomposes into the product of trajectory-level advantage and an entropy sensitivity function, creating an advantage-surprisal four-quadrant structure near criticality. STARE corrects this by reweighting token-level advantages based on surprisal, leaving the reward structure untouched. Teams running RLVR post-training who see entropy collapse mid-run have a targeted diagnostic and fix here. link

03 [Inference] EfficientRollout: System-Aware Self-Speculative Decoding for RL Rollouts A small fraction of long-tailed rollout generations determines the wall-clock time of an entire RL training step, not average-case generation speed. EfficientRollout applies draft-then-verify speculative decoding during rollout sampling to cut that tail latency, using the policy model itself as the drafter rather than a separate smaller model, which avoids introducing a mismatched distribution. The reward model and policy weights stay unchanged. Teams bottlenecked on rollout throughput during RL training can apply this without restructuring the training loop. link

04 [Eval] Beyond the Current Observation: Evaluating Multimodal Large Language Models in Controllable Non-Markov Games Frontier multimodal models fail at acting on observations that are no longer visible, even when their general reasoning otherwise succeeds, and existing benchmarks cannot isolate this failure. RNG-Bench introduces two games, Matching Pairs and a complementary task, where card or object identities are hidden after initial exposure, forcing agents to reconstruct past observations during active multi-step interaction rather than after episode end. This separates hidden-state reconstruction from perception and planning skills that current benchmarks conflate. Teams deploying multimodal agents in partially observable environments should treat this as a distinct capability gap to test for. link

05 [Agent] CEO-Bench: Can Agents Play the Long Game? Agents that score well on short-horizon benchmarks like SWE-bench degrade sharply when required to operate a simulated startup across 500 days under changing world state and noisy information. CEO-Bench bundles four capabilities that isolated benchmarks structurally cannot surface together: long-horizon planning under uncertainty, noisy information acquisition, adaptation to world changes, and orchestration of multiple concurrent objectives. The 500-day simulation horizon forces compounding errors to surface in ways that single-task evaluations absorb and hide. Any team using short-horizon benchmark scores to project agent reliability on real-world deployments should treat those projections as overestimates. link

06 [Theory] Beyond Alignment: Value Diversity as a Collective Property in Multicultural Agent Systems A multi-agent system where every individual agent aligns to its assigned culture can still collapse cultural plurality at the system level if agents converge on shared outputs. Per-agent alignment scores, the current standard for evaluating multicultural AI, are blind to this failure mode because alignment is a per-agent property and cannot measure dissimilarity across agents taken together. This paper formalizes value diversity as a system-level evaluation axis defined through the dissimilarity between culturally conditioned agents' responses, and proposes collective metrics that per-agent scores cannot detect. Teams deploying multicultural agent systems for globally diverse settings need system-level diversity audits, not just per-agent culture-match scores. link