Meta AI AI News, Models and Product Updates

Meta Ai news, model releases, product launches, research updates, and major announcements in one place.

Latest Articles

• LFM2.5 Q4\_0 Checkpoints from Quantization-Aware Distillation — Liquid AI releases QAD Q4_0 GGUF checkpoints for four LFM2.5 models (230M, 350M, 1.2B, 2.6B), recovering ~97% of BF16 accuracy at native 4-bit speed and memory — making high-quality edge inference practical without larger quantization formats.
• Explicit State Elicitation Is Not Enough: A Controlled Audit of Memory-Policy Classification — A controlled audit finds that prompting LLMs to output explicit memory-policy state classifications does not significantly improve routing accuracy for Llama-3.3-70B or GPT-OSS-120B — and that standard example-level accuracy metrics overstate counterfactual consistency.
• DEPT: Document Embedding Preservation Tuning for Unified Query Expansion and Retrieval — DEPT trains a single decoder-only LLM end-to-end as both query expander and retriever, solving the moving-target problem by anchoring document embeddings while retrieval gradients improve query generation — outperforming staged and independent baselines on five BEIR datasets.
• TileMix: Tile-Centric Mixed-Precision Attention for LLM Inference Acceleration — TileMix routes LLM attention tiles through FP16 or INT8 at inference time using compact bitmasks — improving long-context prefill throughput over FP16 while recovering quality lost under uniform INT8, tested on LLaMA, Qwen, and Vicuna with no training required.
• ARENA: Automated Red-Teaming for Large Audio Language Models — ARENA is an automated red-teaming framework for large audio-language models that crafts audio inputs which, combined with safe text queries, induce harmful outputs — achieving attack success rates of 87.9% on Audio Flamingo 3, exposing safety blind spots invisible to text-only testing.
• GraniKV: Asymmetric Granularity KV-Cache Paging for Multi-Agent Systems with Long Shared Prefix — GraniKV separates KV-cache storage into a contiguous pool for shared LLM prefixes and a token-level pool for per-request suffixes, delivering 2.16× output throughput over production baselines for multi-agent workloads with 16K shared prefix tokens on Llama-3.1-8B.
• Improving Influence-based Instruction Tuning Data Selection for Balanced Learning of Diverse Capabilities — BIDS (Balanced and Influential Data Selection) fixes a systematic bias in influence-based fine-tuning data selection—where high-influence tasks crowd out others—and a 15% data subset selected by BIDS outperforms full-dataset training on Llama-3 and Mistral-v0.3 across 7 benchmarks spanning 5 capabilities.
• LlamaRec-LKG-RAG: A Single-Pass, Learnable Knowledge Graph-RAG Framework for LLM-Based Ranking — LlamaRec-LKG-RAG adds a learnable knowledge graph layer to LlamaRec, extracting personalized relation paths from user-item graphs and injecting them into Llama-2 prompts — improving MRR, NDCG, and Recall over the base LlamaRec on ML-100K and Amazon Beauty in a single end-to-end trainable pass.
• Mitigating Rubric Interference in LLM Judges via On-Policy Self-Distillation — SARA (Self-Anchored Rubric Alignment) fixes rubric interference in LLM judges—where multi-rubric batch evaluation produces inconsistent verdicts depending on which criteria co-appear—using on-policy self-distillation with no external supervision required.
• Incoherent by Design? On the Moral Self-Consistency of LLMs — Testing GPT, Mistral, and Llama on morally equivalent scenarios phrased under deontology, utilitarianism, and virtue ethics reveals contradiction rates up to 78% -- meaning LLMs used in ethics-sensitive contexts frequently contradict their own moral positions when context shifts.