PulseAugur / Brief
EN
LIVE 02:34:13

Brief

last 24h
[5/5] 221 sources

Multi-source AI news clustered, deduplicated, and scored 0–100 across authority, cluster strength, headline signal, and time decay.

  1. TOOL · AWS Machine Learning Blog English(EN) ·

    Announcing OpenAI-compatible API support for Amazon SageMaker AI endpoints

    Amazon SageMaker AI now offers OpenAI-compatible API support for its real-time inference endpoints. This integration allows users to invoke models hosted on SageMaker using existing OpenAI SDKs, LangChain, or Strands Agents by simply updating the endpoint URL. The new feature supports bearer token authentication for secure access and enables multi-model hosting and the deployment of fine-tuned open-source models without requiring code modifications. AI

    Announcing OpenAI-compatible API support for Amazon SageMaker AI endpoints

    IMPACT Simplifies integration for developers using OpenAI's ecosystem with models hosted on AWS infrastructure.

  2. TOOL · arXiv cs.CL English(EN) ·

    X-Token: Projection-Guided Cross-Tokenizer Knowledge Distillation

    Researchers have developed X-Token, a novel knowledge distillation technique designed to improve student models by learning from teacher models with different tokenizers. The method addresses limitations in existing logit-based distillation, such as the uncommon-token failure and over-conservative matching, which can suppress critical tokens or exclude near-equivalent ones. X-Token utilizes a sparse projection matrix to align student and teacher distributions, outperforming current state-of-the-art methods on benchmarks like GSM8k and achieving significant gains with multi-teacher setups. AI

    IMPACT Improves cross-tokenizer knowledge transfer, potentially enabling more efficient training of diverse language models.

  3. TOOL · arXiv cs.LG English(EN) ·

    Fine-Tuning Without Forgetting via Loss-Adaptive Learning Rates

    Researchers have developed a new method called FINCH to address catastrophic forgetting during the fine-tuning of large language models. FINCH employs a loss-adaptive learning rate schedule that decreases the learning rate for high-loss batches and increases it as the model converges. This approach effectively reduces forgetting by an average of 93% across various benchmarks while maintaining task performance. FINCH also shows improvements in preserving model calibration and confidence. AI

    Fine-Tuning Without Forgetting via Loss-Adaptive Learning Rates

    IMPACT FINCH significantly reduces catastrophic forgetting in LLMs, potentially enabling more effective and stable fine-tuning for specialized tasks.

  4. TOOL · arXiv cs.CL English(EN) ·

    Internalizing Tool Knowledge in Small Language Models via QLoRA Fine-Tuning

    Researchers have developed a method to internalize tool knowledge into small language models using QLoRA fine-tuning, reducing the need for explicit tool schemas in prompts. By training models like Gemma 4 E4B and Qwen3-4B on tool-use examples, they achieved better planning scores than a baseline that received full tool descriptions. This approach significantly cuts down on input length and inference overhead while maintaining or improving tool-planning quality, though it may impact general knowledge retention. AI

    Internalizing Tool Knowledge in Small Language Models via QLoRA Fine-Tuning

    IMPACT Enables more efficient use of smaller models in agentic systems by reducing prompt token overhead.

  5. RESEARCH · arXiv cs.CL English(EN) · · [5 sources]

    CEPO: RLVR Self-Distillation using Contrastive Evidence Policy Optimization

    Researchers have developed new self-distillation techniques for large language models to improve their performance without relying on external feedback. AVSD (Adaptive-View Self-Distillation) balances consensus signals across multiple privileged information views with view-specific residuals to enhance learning. Self-Policy Distillation (SPD) extracts a capability subspace from gradients to improve performance and generalizability, particularly in code generation and mathematical reasoning. CEPO (Contrastive Evidence Policy Optimization) sharpens credit assignment at decisive tokens by contrasting correct answers with incorrect ones, improving accuracy on multimodal mathematical reasoning benchmarks. AI

    CEPO: RLVR Self-Distillation using Contrastive Evidence Policy Optimization

    IMPACT These self-distillation techniques offer improved performance and generalizability for LLMs in complex reasoning tasks without external supervision.