Core ML efficiency/interpretability (Recurrent TF, DeepSeek-V4, ShadowPEFT)

Diffusion spec 3x; Meta OMT-LLaMA; SGS 7B>671B; Signpost/Hier mem/RoundPipe/DORA/LenVM/LeCun JEPA; MSA 100M mem; efficient-transformers LVLM mem issues. New: Flow map language models (FMLM) outperform discrete diffusion in few-step generation; adversarial flow distillation for video; OCTOPUS extreme KV cache compression (near-optimal, zero latency); RT-Lynx activation sparsity for diffusion transformers (1.55x speedup); EDGE-OPD on-policy distillation; MEMO modular memory model for continual learning (outperforms RAG); Stanford HAI scaling law method for one-shot LLM training. Also: Gemini Embedding 2 native multimodal embedding (SOTA on multiple benchmarks); AKBE efficient agentic RL; SAM state-adaptive memory. Non-AR boosts accelerate. New: Queue & AI paper introduces 'variance wedge' concept for AI workflow slowdowns — relevant to deployment efficiency. New: DenoiseRL efficient RL from noisy prefixes; BES bidirectional evolutionary search for self-improvement. New: How LoRA Remembers? (parametric memory law for LoRA, phase transition at p>0.5, MemFT optimization). New: Thinking Before Constraining (unified decoding framework, up to 27% accuracy gains over natural generation); Why Larger Models Learn More (mechanistic explanation: capacity, interference, rare-task retention, validated with OLMo 4M-4B). New: CollectionLoRA (distilling 50 effect LoRAs into one via multi-teacher on-policy distillation, addresses parameter interference and deployment overhead). New: DiffusionBlocks (Sakana AI) — block-wise training via diffusion, new paradigm for scalable neural network training. New today: Trajectory concurrent multi-LoRA training stack (2.81× throughput gain, no accuracy regression, vLLM multi-LoRA inference). Also: LoopMDM (looped diffusion language models, depth-scaling without extra parameters, flexible compute scaling). Also: SAERL (SAE-guided RL post-training data engineering, +3% accuracy, -20% training steps) adds data-centric efficiency. New: dMoE (dLLMs with learnable block experts, reduces activated experts from 69.5 to 14.6, 99% performance). New: Learn from your own latents (theoretical proof that latent prediction beats token prediction for sample complexity, Random Hierarchy Model). Also: TurboQuant open-source 5x KV cache memory reduction (Tether AI, production-ready). New: MiniMax-M3 sparse-attention model beats GPT-5.5 and Gemini 3.1 Pro on agent benchmarks at 5-10% cost — major efficiency breakthrough. Also conceptual signal: @jon_barron argues autoregression and diffusion are not a false dichotomy, citing diffusion forcing. New articles: On the Scaling of PEFT (million personal models, MinT infrastructure for adapter management), NITP (Next Implicit Token Prediction, +5.7% MMLU-Pro, low-overhead pre-training improvement), ESPO (early-stopping PPO, >20% token savings, improves math reasoning). New: Agentix (NSDI '26) — efficient serving engine for LLM agents as general programs, 4-15x throughput gain over vLLM by treating programs as first-class citizens and preempting calls based on program context. New: Oryx — dynamic attention & recurrent hybrid LLM, ties 90% of parameters, balancing long-context retrieval and generation speed. New: GPU kernel performance forecasting via calibrated LLM surrogates (12k kernel dataset, selective prediction, enables efficient evolutionary search for self-improving agents). New: VaSE (value-aware stochastic KV cache eviction for reasoning models, 4x compression, training-free). New: Decoupled Residual Denoising Diffusion Models (unified I2I translation, data-efficient, CVPR). New: HyLo hybrid architecture (upcycles existing LLMs with MLA+linear blocks, 90% KV-cache reduction, 32x context extension). New: TrOPD (stable on-policy distillation for reasoning models, partitions tokens into trust regions and outliers). New: Gemma 4 12B (encoder-free multimodal, laptop-ready, near 26B MoE performance) — notable release. New: NVIDIA Nemotron 3 Ultra — hybrid Mamba-Transformer, NVFP4 quantization, LatentMoE, multi-token prediction, Multi-Teacher On-Policy Distillation, 5x throughput on agentic tasks. New signal: @mmbronstein reposted MUX — latent continuous reasoning from text CoT, new direction for reasoning efficiency. New: NF-CoT (latent reasoning with normalizing flows) preserves KV-cache and tractable likelihood, improves code-gen pass rates at lower cost. New: OPRD (On-Policy Representation Distillation) eliminates sampling variance, 1.44x speedup, 54% memory reduction. New: Fine-tuning vs ICL study (ACL 2026) — fine-tuning converges to equal proficiency, ICL shows variability; both generalize similarly OOD. New: MLEvolve (self-evolving framework for automated ML algorithm discovery, SOTA on MLE-Bench) — relevant to core ML automation. New: AdaCodec (predictive visual code for video MLLMs, 1/7 token budget, 5.7x TTFT reduction). New: Flash-WAM (modality-aware distillation for world action models, 23x speedup). New: Video2LoRA (parametric video internalization, 1500x token reduction). New: Combinatorial Synthesis (atomic decomposition for code RLVR, scaling data generation). Also: AntiSD (reverse self-distillation, 11.5% math gains, 2-10x fewer steps), Do Transformers Really Need Three Projections? (K/V sharing, 50% cache reduction, <3% perplexity loss). New: Stateful Encoders (VLMs with visual memory via cross-attention and stop-gradient, consistent gains across backbones, practical for multi-image reasoning). New: New Sleep Paradigm for LLM Memory Consolidation (two-stage sleep with distillation and dreaming, biologically inspired, addresses forgetting). New: Compress-Distill (reasoning trace compression, 7.6x training speedup, 96% accuracy retention). New: ADASPEC (multilingual speculative decoding, 2.3x speedup over EAGLE-2, self-synthesized training data). New: On-policy self-distillation conditioning on own outputs outperforms vanilla on-policy RL (tweet). New: Forgetting in LMs — self-generated replay with BOS token reduces forgetting to near zero, 38x efficiency. New today: LCLMs (end-to-end context compression at scale, targets long-horizon agents, architecture search, 350B token pre-training), Reasoning Arena (trace tournaments for RLVR, 7.6% gain, 27-41% speedup), FNS tokenizer (hybrid subword-character resolution via loss augmentation, practical for agentic workflows). New: DRPO (smooth quadratic divergence regularization for LLM RL), FlowTracer (attention-induced information flow for token-level credit assignment, ICML 2026). New: Harness-1 (20B state-externalizing search agent) also contributes to efficiency via reduced context overhead. New: DiffusionGemma (4x faster text generation, open diffusion model) — significant speedup for text generation. New: Redesign MoE routers with manifold power iteration — potentially impactful for efficiency. New: Verifiable Environments as LEGO bricks — recursive composition for reasoning generalization. New today: Adaptive Multi-Resolution Procedural Knowledge Compression for LLMs — multi-resolution compression for flexible compute-memory trade-offs. New from today's articles: HarnessBridge (bidirectional controller for token efficiency), SG-OPD (sign-gated on-policy distillation), MaxProof (generative-verifier RL for math proof scaling). Also new: Nightjar (dynamic adaptive speculative decoding, addresses KV cache overhead under high load). New: N-GRPO (embedding-level neighbor mixing for exploration in GRPO, improves math reasoning). New: Demystifying Hidden-State Recurrence (Switchable Latent with Switch-GRPO, latent computation with curriculum learning). New: SMEPilot CPU inference optimization (ex-6cff8086); FastContext 4B repo exploration model (ex-1496e78c). New today: @StanfordHAI tweet on new scaling law method to reduce training demands (ex-1099a237) — relevant to efficiency. New: Only Loop Once for efficient test-time computation scaling (reduces latency/KV-cache in Looped Transformers). New: VibeThinker-3B small model matches DeepSeek V3.2 and Gemini 3 Pro on reasoning (Parametric Compression-Coverage Hypothesis). New: @jeremyphoward repost signals return of critics (value functions) in RL training for long-horizon tasks, challenging GRPO. New: VibeThinker-3B also matches Opus 4.5 in programming — reinforces small-model reasoning trend. New: How Efficient Attention Shapes Hybrid LLMs (ex-f25db1d6) — Large-Window Laziness and NoPE boost long-context performance. New: Explaining Attention with Program Synthesis (ex-36f308c5) — symbolic interpretability method. New: V-Zero (answer-label-free on-policy distillation for visual reasoning, 5-10x speedup), CAVEWOMAN (input compression backfires, output compression saves cost). New: JetSpec (parallel tree speculative decoding, 9.64x speedup, breaks scaling ceiling). New: Error-Conditioned Neural Solvers (10× gain on turbulent Kolmogorov flow) adds core physics-informed ML improvement. New practical observation: @rasbt reports local 30B MoE models (Qwen-Code, Codex, Claude Code) hitting ~40 tok/sec on Mac/DGX Spark, matching GPT 5.5 subscription speed; Claude Code uses 2x tokens vs Codex, highlighting harness efficiency differences. New: ELDR (expert-locality-aware decode routing for PD-disaggregated MoE serving, 5.9-13.9% TPOT reduction) adds practical MoE efficiency. New: Multimodal Continuous Reasoning via AMVL (asymmetric mutual variational learning, +10.83 on BLINK) improves multimodal reasoning efficiency. New today: low-bit quantization inflates reasoning tokens (CoT Token Inflation Ratio, QAT helps, mitigations not solved). New: HOLA (compressive recurrent state + small exact cache, sub-full-attention perplexity, 16x length extrapolation) adds practical efficiency. New: Survey on efficient attention mechanisms provides overview. New: 'meek models' paper on bounded vs unbounded metrics challenges democratization narratives. New: OrbitQuant (data-agnostic quantization for DiTs, W2A4 usable quality, cross-modality transfer). New: MIPI/MIPU (monotonic inference policies as real objective for LLM RL, addresses training-inference mismatch). New: OmniOpt (comprehensive optimizer survey and benchmark) adds reference for training efficiency. New: Carmack's NAND flash for AI inference (hardware-level cost argument, deterministic access patterns) adds provocative system-level efficiency perspective. New: Local LLM energy benchmark (3-shot prompting improves efficiency without accuracy loss) adds practical deployment insight. New today: HiLS Attention (hierarchical sparse attention, 64x length extrapolation, 90% retrieval accuracy) adds practical long-context efficiency.

Status: Climaxing — efficiency breakthroughs from MiniMax-M3, Nemotron 3 Ultra, and various KV cache/attention optimizations. The return of critics and value functions challenges GRPO dominance. Small models (VibeThinker-3B) matching frontier models on reasoning is a notable trend. The energy cost paper (H8) adds system-level perspective.