Also Worth Noting - 2026-08-20

Also Worth Noting

02 [Training] Co-RL: Unsupervised Reasoning Emerges from Diverse Cohort in Multi-agent RL Ground-truth annotations are not required for RL-based reasoning gains. Co-RL uses disagreement across a cohort of agents as the reward signal, bypassing the need for verifiable answers entirely. Where self-rewarding methods reinforce existing model biases by looping on their own outputs, cohort disagreement surfaces genuine uncertainty and drives improvement. Teams working in domains where correct answers are unverifiable or too costly to label should watch this closely. link

03 [Training] The More Popular, The Harder to Forget: Adaptive Popularity for LLM Unlearning Flat gradient pressure for unlearning is a compliance risk: facts memorized more frequently during pretraining resist removal far longer than rare ones, meaning the highest-risk information survives most unlearning runs. AdaPop assigns a popularity-dependent exponent to each fact, derived from an external proxy such as Wikidata sitelinks or LLM-as-Judge scoring, and automates the forget-retain tradeoff via a dual-ascent controller that adjusts the retain penalty each epoch. Across three model families and two benchmarks, AdaPop leaks roughly 5x less memorized content. Teams running compliance-driven unlearning should audit whether their budget scales with training frequency, not just fact count. link

04 [Open-source] Training Leaves Traces: Centered Residual Signatures for Language Model Lineage Verification A shared identity-aligned component embedded in residual block weights survives fine-tuning, quantization, pruning, and merging, making weight-only ancestry detection possible without any training data. Removing that shared component and comparing checkpoint-specific residual structure across blocks yields a symmetric lineage score that separates related from unrelated checkpoints on residual-MLP and GPT-2 benchmarks. No model cards, no data access, no metadata required. Provenance audits of open-weight releases now have a concrete, data-free path forward. link

05 [Agent] Looped Language Models Improve Compositional Tool Calling Recurrence, not chain-of-thought prompting, is what multi-step API coordination actually needs. Looped language models reuse intermediate hidden states across loop iterations rather than collapsing everything into a single forward pass, letting the model maintain dependencies across tool calls without external scaffolding. Evaluated on API-Bank, BFCL, and NESTful under matched supervised fine-tuning recipes, looped models consistently outperform their non-looped counterparts on compositional tasks. Teams building agents that chain more than two or three tool calls should treat recurrent depth as a tunable inference-time variable, not a fixed architectural afterthought. link

06 [Inference] ReCache: Efficient KV Cache Reuse and Compression for Tool-Augmented LLM Agents Standard prefix caching breaks the moment tool schema order varies across requests, which is almost always in production agentic workloads. ReCache solves this by caching KV states per resource independently, assigning resource-local positions and removing cross-resource attention interactions to produce composition-invariant KV blocks. At inference time, blocks reorder to match the current request without recomputation, recovering prefix-cache savings regardless of schema shuffle. Teams running high-QPS agentic APIs get memory and latency reductions without any retraining. link