LLMs

DBLAST addresses a critical flaw in block diffusion drafters where the assumption of conditional independence between draft positions breaks down during non-greedy, stochastic sampling. The authors demonstrate that as the entropy of the target distribution increases, the length of accepted drafts degrades significantly. This method aims to realign the drafter's training with the stochastic realities of production inference.

• Standard block drafters assume independent positions, causing failure under stochastic sampling.
• Accepted draft length degrades as target distribution entropy increases.
• DBLAST introduces dependent block drafting to handle multiple plausible continuations.
• Crucial for optimizing speculative decoding in non-greedy LLM inference scenarios.

This 2025 article provides a deep technical breakdown of vLLM's architecture, focusing on how it achieves high throughput for large language model inference. It details core components such as PagedAttention and continuous batching, which optimize GPU memory usage and handling of variable request lengths. The piece serves as a comprehensive guide for engineers seeking to understand the underlying mechanics of this popular open-source inference engine.

• PagedAttention solves memory fragmentation issues inherent in traditional KV cache management.
• Continuous batching strategies significantly improve GPU utilization by handling variable request lengths.
• The architecture is designed for efficient scaling on modern GPU hardware.
• Provides a clear view of the internal mechanics driving vLLM's performance advantages.

SemiAdapt-Instruct addresses the difficulty of updating instruction-tuned LLMs for new domains without full retraining. The framework uses a modular approach to discover latent instruction domains and trains parallel LoRA adapters for each. It employs parameter-free routing to combine these adapters, allowing new domains to be added by training only a single new adapter.

• Enables adding new domain capabilities without retraining the entire model or existing adapters.
• Uses parameter-free routing to combine outputs from multiple parallel LoRA adapters.
• Outperforms full model fine-tuning on ROUGE-L and LLM-as-a-judge metrics.
• Matches single LoRA fine-tuning performance while offering superior extensibility.
• Discovers latent instruction domains automatically to structure the adapter training.

A study on 24,000 judgments reveals that forcing models to persist evidence in separate API calls before rendering a verdict reduces alignment with human preferences by 4-6 percentage points. The research compares standard pairwise judging against structured single-call and multi-call evidence locking workflows using Claude Sonnet 4.5 and GPT-5. Results indicate that isolating evidence extraction from the final decision-making step negatively impacts evaluation quality.

• Evidence locking reduces human preference agreement by 4-6 percentage points compared to standard judging.
• Multi-call workflows (2-3 calls) performed worse than standard pairwise or single-call structured judging.
• Isolating evidence extraction from the verdict step degrades evaluation quality for reasoning models.
• Standard pairwise or single-call approaches may yield more reliable LLM-as-Judge outcomes.

Research demonstrates that large language models often fail reasoning tasks due to localized intermediate errors rather than global incompetence. A weak probe model can generate short patches to correct these bugs and redirect the strong model's trajectory toward the right answer. However, directly fine-tuning on these repairs fails to internalize the fix, indicating the value lies in reshaping the reasoning distribution rather than the patch text itself.

• Reasoning failures are often localized bugs fixable via weak-model patches.
• Direct fine-tuning on repaired trajectories does not reliably internalize corrections.
• The key signal is how patches reshape the model's future reasoning distribution.
• Weak models can effectively diagnose and patch strong model reasoning errors.

This research introduces a method where procedural scaffolds are organized into an evolvable graph that co-evolves with model parameters during post-training. By integrating discovery, distillation, and dynamic recompilation, the approach addresses the disconnect between parameter optimization and independent inference-time strategies. The resulting Skill Training paradigm demonstrates significant gains on FeatureBench, with auto-discovered skills improving pass rates by 8.1 percentage points.

• Co-evolution of parameters and procedural graphs enables automatic acquisition of complex strategies.
• Skill Training improves FeatureBench pass rates by 8.1pp via automatically discovered skills.
• Progressive distillation retains 85.2% of performance gains without external scaffolds.
• Models achieve a 27.7% pass rate post-distillation, proving effective knowledge internalization.

This study audits PII leakage in English-source multilingual RAG systems using Qwen2.5-7B across five query languages. It challenges the assumption that non-English queries are inherently safer, finding that English exhibits the highest unstructured PII leak rate under output-only filtering. Adding an input judge shifts residual leak risks to Arabic and Swahili, highlighting that pipeline architecture dictates risk more than language choice.

• Non-English queries are not automatically safer; risk depends on pipeline stages.
• Output-only filtering fails hardest for English PII in this specific setup.
• Input judges shift residual leaks to Arabic and Swahili rather than eliminating them.
• Findings are specific to the Qwen2.5-7B pipeline, not universal language traits.
• Defenses must be evaluated across the full translation-judge-generation chain.

DREAM introduces an Event-aware Memory Graph (EMG) to solve the long-term coherence issues in LLM-based role-playing agents. Inspired by the ABC cognitive model, it transforms unstructured literary text into a structured graph of events, beliefs, and consequences. This approach moves beyond static character descriptions to enable causally grounded behavioral reasoning over time.

• Replaces unstructured memory with a graph structure for better narrative consistency.
• Uses the ABC model to link events to character beliefs and consequences.
• Enables long-term personality coherence in immersive role-playing simulations.
• Transforms literary text into structured data for causal reasoning.
• Addresses key limitations of existing static role-playing agent frameworks.