Every processed story in chronological order, with the newest coverage first. Filter by tag, source, or score to drill in.
InterleaveThinker removes the architectural barrier that has prevented existing image generators from producing interleaved text-image sequences, extending a capability previously limited to frontier models like GPT-5 to any image generator via a plug-in multi-agent pipeline.
The experiment demonstrates that Haiku 4.5's tendency to honestly acknowledge logical inconsistencies — while a virtue in cooperative contexts — made its negotiating position progressively indefensible against an adversarial attacker, in contrast to Opus 4.8's strategy of holding a single, unreinterpreted constraint throughout.
HORMA reduces agent token consumption to at most 22.17% of baseline while maintaining or improving task performance, directly addressing the inference cost and latency penalties that make long-horizon LLM agents expensive to run.
The paper demonstrates that fabricated success in unattended LLM agents is a structural problem solvable by gate enforcement rather than model selection, reducing SWE-bench Lite fabrication by over 33 percentage points compared to the StateFlow baseline.
The tutorial makes the architecture of Gated DeltaNet fully derivable from first principles by tracing the exact theoretical lineage through linear attention, SSMs, and Mamba, rather than presenting it as a black-box model.
The results show that targeted RL fine-tuning on high-quality, task-specific data can close — and reverse — a 231-billion-parameter gap in model size, at a training cost under $500, on a real financial reasoning benchmark.
The results show that structured verifier feedback — not just more data — can unlock large performance gains for LLM agents on formal reasoning tasks, pointing toward a concrete path for verifier-guided program synthesis.
AutoPDE's explicit strategy representation closes a key gap in LLM-based PDE solvers, where numerical decisions previously remained hidden in code and were difficult to inspect or correct when solves failed.
HIPIF directly targets long-context interference — a problem existing hierarchical RL and credit-assignment methods leave unaddressed — by folding completed subgoal histories, offering a path to more reliable LLM agent performance on extended, multi-turn tasks.
The post reports that Fable 5 tops coding and reasoning benchmarks and delivered immediate, measurable acceleration on large-scale real-world tasks, marking a notable step-change in agentic coding capability.