LLMs

Existing position-independent caching (PIC) methods rely on token-indexed KV caches, which do not exist in hybrid LLMs that use linear recurrences. LinearKV introduces a training-free framework that enables PIC for these models by decoupling the initialization of the fixed-size state. This approach allows hybrid models to reuse token chunks and restore context without the standard KV concatenation primitives.

• Hybrid LLMs lack token-indexed KV caches, breaking standard PIC methods.
• LinearKV is a training-free framework designed specifically for hybrid architectures.
• It uses decoupled initialization to manage fixed-size linear recurrence states.
• Enables caching benefits for models using linear attention layers.
• No need to rebuild entire context when reusing token chunks.

The DeepSeek V4 Pro 0813 model is now available on OpenRouter, with technical specifications and API documentation hosted on the official DeepSeek developer portal. Artificial Analysis provides performance benchmarks for this iteration, while community discussions on Hacker News highlight technical observations and usage patterns.

• Access V4 Pro 0813 through OpenRouter's standardized API endpoint.
• Review official API docs for updated rate limits and parameter schemas.
• Check Artificial Analysis for latency and throughput benchmarks.
• Monitor HN threads for real-world inference performance reports.

Switchyard is an open-source Rust library and proxy designed to manage large language model traffic. It translates between OpenAI, Anthropic, and OpenAI Responses API formats, allowing clients to interact with backend providers like vLLM, NVIDIA NIM, or Ollama without code changes. The tool supports advanced routing strategies, including A/B benchmarking and signal-driven stage routing, while recording operational metrics.

• Use Switchyard to decouple coding agents from specific LLM providers via API translation.
• Route traffic across multiple models (vLLM, NIM, Ollama) for A/B testing and benchmarks.
• Implement custom or signal-driven routing algorithms with typed, composable logic.
• Capture operational metrics for LLM traffic without modifying client code.

Research reveals that context compaction mechanisms in LLMs frequently discard user-issued session constraints, such as specific behavioral rules or data retention instructions. The new COMPINT evaluation suite shows current compactors retain only 17% of these constraints on average, often performing worse than tasks without compaction. This loss occurs across multi-turn chats, agentic trajectories, and long-horizon research scenarios.

• Session Constraints (SCs) are silently dropped during context compaction, breaking user intent.
• COMPINT suite evaluates compactor performance across chat, agentic, and research tasks.
• Average SC retention is a mere 17%, with many compactors underperforming non-compacted runs.
• Retention rates vary significantly based on compactor model, prompt style, and context length.

Researchers introduce Self-Fix Step-DPO (SFS-DPO), a two-stage reinforcement learning framework designed to improve how large language models verify and correct their own errors. The first stage optimizes step-level reasoning through preference optimization, while the second explicitly trains the model for self-verification and correction. A teacher-assisted variant, SFS-DPO-R, further enhances this by incorporating explanatory rationales to provide stronger corrective signals during the training process.

• Two-stage RL approach targets step-level reasoning rather than just final output correctness.
• SFS-DPO-R uses teacher-generated rationales to strengthen error verification signals.
• Framework demonstrates effectiveness across both in-domain and out-of-domain LLM evaluations.
• Addresses the core challenge of enabling LLMs to reliably self-correct mistakes.

Researchers address the composability gap in unstructured knowledge editing, where current models can recall injected passages but fail to use them for multi-hop reasoning or atomic fact retrieval. The proposed method employs a hybrid-policy self-editing mechanism to ensure the model actually leverages the new information. This approach allows LLMs to update specific knowledge without degrading unrelated capabilities, enabling true composability of unstructured facts.

• Current unstructured editors fail at multi-hop reasoning despite successful recall
• New hybrid-policy method ensures edited knowledge is actively usable
• Enables composability of free-form passages without catastrophic forgetting
• Addresses static training data limitations for fast-changing knowledge domains