LLMs

CriticGen replaces coarse, decoupled LLM evaluation with a framework that generates sample-specific criteria dynamically. It produces a score, reasoning, and executable refinement suggestions based on these custom rubrics. This process enables models to diagnose specific flaws and immediately apply improvements to their own outputs.

• Moves beyond static scoring by creating dynamic, sample-specific evaluation rubrics.
• Outputs executable refinement suggestions rather than generic critique.
• Enables closed-loop model improvement through generation-aware feedback.
• Categorizes constraints into subjective, objective, and self-derived types.

Researchers propose Neuron-Guided Fine-Tuning (NGFT) to address parameter redundancy and catastrophic forgetting in large language models. The framework uses neuron activation patterns as a universal proxy to synchronize data selection, parameter updates, and knowledge preservation. It identifies essential neurons in a single forward pass to concentrate updates and reduce computational waste.

• NGFT unifies data selection and parameter updates using a single optimization signal based on neuron activations.
• Adaptive Task-Specific Neuron Selection reduces redundancy by focusing updates only on essential neurons.
• The method mitigates catastrophic forgetting by preserving knowledge through activation-based constraints.
• Serves as a holistic alternative to isolated SFT improvements like LoRA or standard data filtering.

AWS has released GPT-6 Astra from OpenAI to general availability on Amazon Bedrock. This model features a 1 million input token context window and enhanced reasoning capabilities for complex business workflows. It integrates with Amazon Bedrock APIs and supports ChatGPT Work and Codex configurations.

• GPT-6 Astra supports up to 1 million input tokens for large context analysis.
• Available via Amazon Bedrock APIs for production-grade inference and security.
• Integrates with ChatGPT Work and Codex on Amazon Bedrock.
• Optimized for autonomous agents and professional-quality content generation.

This paper introduces a compression framework for Large Reasoning Models that avoids uniform INT4 quantization, which can damage critical reasoning circuits. By profiling per-module vulnerability via perturbation sweeps, the method selectively restores sensitive components to FP16. Experiments across five reasoning benchmarks reveal that aggressive quantization can paradoxically increase energy usage by extending reasoning chains.

• Uniform INT4 quantization risks damaging critical reasoning circuits in LRMs.
• The framework profiles vulnerability across 196-224 layer/projection pairs.
• Sensitive circuits are selectively restored to FP16 to preserve accuracy.
• INT4 quantization may increase energy by forcing longer reasoning chains.
• Energy efficiency requires balancing compression with reasoning integrity.

Researchers introduce Intra-Prompt Parallel Decoding (IPPD) to answer multiple questions sharing a context within a single prompt. This method bypasses the memory bottleneck of separate prompts by sharing attention computation and memory. The approach directly targets GPU underutilization in common-context question answering tasks.

• IPPP packs multiple questions into one prompt to share attention overhead.
• Reduces memory bottlenecks inherent in independent prompt batching.
• Improves GPU utilization for common-context QA workloads.
• Offers a new inference strategy beyond standard KV-cache techniques.

Research indicates that large language models can generate new social biases not present in their training data through the process of adaptive exploration. This suggests that the emergent behaviors during inference or fine-tuning may diverge from expected safety boundaries. The findings highlight a potential risk where models autonomously construct prejudiced patterns while optimizing for performance or novelty.

• Bias emergence is not limited to training data leakage; it can arise dynamically.
• Adaptive exploration mechanisms may inadvertently foster novel social prejudices.
• Standard safety filters might miss biases that are generated post-training.
• Monitoring inference-time behavior is critical for detecting emergent risks.
• Current alignment techniques may need updates to address autonomous bias formation.

Recent benchmarks of the Qwen3.8 27B model reveal a sharp performance cliff when aggressively quantizing to 1-bit. While 4-bit quantization maintains acceptable accuracy and efficiency, 1-bit representations suffer from significant degradation. The results suggest that 4-bit remains the practical lower bound for this architecture in production workloads.

• 4-bit quantization preserves model quality for Qwen3.8 27B, making it a safe deployment choice.
• 1-bit quantization causes severe accuracy collapse and should be avoided for this model.
• Benchmarks highlight the importance of testing extreme quantization levels before fleet rollout.
• Consider 4-bit as the minimum viable compression ratio for this specific 27B parameter model.