Also Worth Noting - 2026-08-25

Also Worth Noting

02 [Inference] Quantization-Aware Healing: A Practical Recipe for Recovering Compressed, 4-Bit LLMs Standard QAT applied to a structurally compressed, 4-bit model converges slowly and collapses past its own peak. Quantization-Aware Healing sidesteps this by targeting recovery at the specific capability gaps , reasoning, math, coding, and long-context behavior , rather than re-fitting the full compressed model to hard labels. The compute cost drops substantially while recovering most of the quality lost to joint compression and quantization. Teams shipping quantized models at scale should treat this as the default recovery playbook before QAT. link

03 [Training] Beyond Imitation: Filtering On-Policy Distillation by Reasoning Progress Teacher-derived rewards in on-policy distillation routinely conflict with actual reasoning progress: a trajectory step can show clear logical advancement and still receive low teacher reward. Filtering student-generated trajectories by measurable reasoning progress rather than raw teacher signal breaks this mismatch. The result is a student that improves more on math and coding tasks than one trained on unfiltered teacher feedback. Teams running knowledge distillation for reasoning models should audit what their teacher reward is actually selecting for. link

04 [RAG] Better Retrieval, Worse Robustness: How Multi-hop RAG Amplifies Upstream ASR Errors Entity-graph linking and iterative reformulation, two popular upgrades to standard RAG, both amplify ASR transcription errors rather than absorbing them across four English accents and three multi-hop QA benchmarks. Structurally richer retrieval configurations gave better clean-text performance but wider degradation curves once ASR noise entered the pipeline. The failure is upstream and fixed: the model never sees the correct query. Anyone building voice-first RAG products needs to stress-test against accent-varied ASR output before shipping multi-hop configurations. link

05 [Eval] One Polluted Page Is Enough: Evaluating Web Content Pollution in Generative Recommenders A single crafted promotional page in a live web index is sufficient to flip search-augmented LLM recommendations toward fake products. The FORGE benchmark measures this attack surface under controlled pollution conditions, and the finding holds without any model access , only content publication is required. That makes this a live deployment risk, not a theoretical one. Teams shipping generative recommenders over live web retrieval should treat adversarial content injection as a production threat category, not an academic edge case. link

06 [Inference] TileMix: Tile-Centric Mixed-Precision Attention for LLM Inference Acceleration Assigning precision at the hardware score-tile level rather than per token or per layer lets TileMix get most of the throughput of INT4 attention while keeping accuracy close to FP16 on long-context prefill. The kernel routes numerical precision as a spatial decision over hardware-aligned tile groups inside fused dense attention, without the token-selection overhead of sparse methods. The gains concentrate exactly where they matter most: prefill at 100k+ token context lengths. Infrastructure teams running long-context workloads should benchmark this against uniform low-precision attention before committing to either extreme. link