Research Briefing

Iris trains two large-scale search agents through a multi-stage pipeline combining supervised fine-tuning and reinforcement learning against live search APIs. It achieves state-of-the-art open-source results on complex web benchmarks via trajectory filtering and inference-time context management.

The paper provides a mechanistic interpretation of chain-of-thought reasoning, showing that distinct reasoning operations are geometrically separable in hidden representations. The structure emerges across layers and depends on contextual reasoning context.

GAPO (Group Adaptive Clipping Policy Optimization) adaptively adjusts importance-sampling clipping thresholds based on rollout advantage in RLVR, preserving stronger gradient signals from low-success groups. It is presented as an improvement over GRPO.

τ^τ-Bench is a benchmark for evaluating coding agents on building real-world customer-service agents from business records, client requirements, and production APIs. The authors find substantial gaps between agent and expert performance.

Continuing our coverage from yesterday, Summarizes OpenAI Chief Scientist Jakub Pachocki's recent public statements on superintelligence timelines, recursive self-improvement, and the lack of preparedness, with commentary on monitorability of the recently released Astra model.

RISE improves LLM post-training by recursively generating dense token-level supervision from the model's own RL trajectories via self-extrapolation, avoiding external teacher models. It fits within the on-policy distillation and RLVR paradigms.

The paper revisits layer dropout (stochastic depth) for LLM training, showing it improves training efficiency and enables faster inference through early exit and speculative decoding without sacrificing accuracy. Authors are from Cerebras, lending industrial relevance.

The paper decomposes safety-tuning responses into refusal statements and rationales, finding that training on rationales alone reduces false refusals while preserving safety. This provides a structural lens on safety-tuning for language models.

Reports a failed replication of Anthropic's J-Lens on GPT-2 small and medium across five stress tests (more data, frequency checks, sparsity, parameter tuning, scale), finding J-Lens loses to Logit Lens at nearly every layer.

The paper formalizes multi-agent LLM coordination as a bilevel game with stochastic memory reflection, introducing a grounded evaluation gate and the SRMA algorithm with convergence guarantees. It validates the approach on SWE-bench, a standard software engineering benchmark.