Google shipped Gemini 4 Neon and DiffuSpace published the technical report for Clover-35B-A3B, an open-weight diffusion language model. OpenAI reports its internal coding agents now complete 46% of 2–4 hour research tasks, while the Resolution Hypothesis paper finds self-evolving agents stall at telling real improvements apart. On efficiency, Late Entry, Chorus and FlashMorph2 cut inference and memory costs; on voice, AssemblyAI released a voice-agent STT model with 86 ms finalization latency.
Large Language Models
1
Late Entry in Looped Language Models
Looped language models rerun a shared layer stack several times per token, which makes decoding slower as depth grows. Instead of early exit, which leaves the KV cache of past tokens shallow, the authors let new tokens skip early passes and join only the final K passes over a deep cache (late entry). After light fine-tuning to fix a hidden-state norm mismatch, it beats matched early-exit decoding across model families. A Wake verification step with a logit-margin tolerance τ unifies greedy-exact, margin-gated and lossy decoding in one static CUDA graph. Ouro-2.6B gets 5.2x greedy-exact speedup and Huginn-3.5B up to 8.4x at T = 64, or 10.6–12.5x lossy.
Why it mattersLooped models are a promising way to buy reasoning depth without more parameters, but their decoding cost has been the main obstacle. Exact speedups of 5–8x without changing outputs make them far more practical to serve.
Chorus: Shared N-gram Embeddings with Layer-wise Adaptation
N-gram embedding tables add capacity cheaply, but per-layer tables are memory-heavy and a single shared table performs poorly. Chorus uses one shared set of n-gram tables for the whole network, with each layer rescaling components based on its hidden state before its own projection. It matches existing methods with 87% smaller tables and outperforms Engram, TN-gram and FactorEngram at table sizes from 20M to 9.6B parameters. The Ember optimizer cuts table-optimizer memory by 99.2% versus AdamW without hurting training.
Why it mattersLookup-table memory is becoming a serious design axis next to MoE as labs look for cheap capacity. Making those tables 87% smaller removes much of their memory cost and makes the idea easier to adopt at scale.
Converting a pretrained Transformer into a hybrid of full and linear attention requires choosing which layers keep full attention. The authors identify two mismatches in the earlier FlashMorph method and propose FlashMorph2, which evaluates sampled hybrid architectures on their own hidden states and jointly optimizes layer-selection scores with model weights. Across Qwen3 models and linear-attention variants it improves on FlashMorph on long-context and recall benchmarks, using only a few GPU hours for layer selection.
Why it mattersConverting existing models is the cheapest route to long-context efficiency, since pretraining hybrids from scratch is expensive. A few GPU hours for layer selection puts this within reach of most teams with open weights.
The Resolution Hypothesis: Agent Evolution's Key Is Identifying Improvements
The paper studies the common agent self-improvement loop in which a proposer generates prompts, skills or memories and a selection step keeps those that score higher. It argues evolution stalls not because proposals are poor but because noisy evaluation cannot tell good proposals from bad. Across seven methods, six models and over 10 million LLM calls, helpful artifacts exist but selection is no better than chance at separating them from harmful ones, so errors compound. Typical gaps between methods are smaller than run-to-run variance, so published rankings may partly reflect luck.
Why it mattersSelf-evolving agents are one of the most active research areas, and this paper suggests much of the reported progress may be noise. It redirects effort toward better evaluation and statistics rather than ever more elaborate proposal schemes.
OpenAI's second report on how coding agents affect its internal AI research, following a September 2026 report. Estimated autonomous success on 2–4 hour tasks rose from 23% in March to 46% in October, and on 4–8 hour tasks from 13% to 35%. Success is jagged: 79% for monitoring training runs, 35% for launching them and 25% for inference and serving code. Agent workdays per researcher rose from 0.5 to 5.7 and median daily spend reached $2,400. In a survey of 101 staff, the median researcher would trade 40% of team headcount or 30% of compute to keep frontier coding agents for three months.
Why it mattersIt is rare first-party data on how much AI is already speeding up frontier AI research, which bears directly on questions about recursive self-improvement. The uneven success across task types shows where human researchers still matter most.
Thinking Fast, Anywhere: One System-1 for the Decision Points of Any Agent Workflow
AnyS1 turns human-written procedural skills into decision points in an agent workflow. The authors mine over 1,000 public skills and train a calibrated 4B System-1 pointer model that picks the next step at each decision point. Routine decisions are resolved in one forward pass, with fallback to the large System-2 model when confidence is low. Across three unseen benchmarks and different System-2 families, it cuts episode cost by 75–91% and model-call time by 63–92% while preserving task reliability.
Why it mattersMost agent steps are routine, yet today each one costs a frontier-model call. A small calibrated router that handles the easy steps is a practical way to make agents several times cheaper without retraining the big model.
Google released Gemini 4 Neon, which sits between Gemini 4 Hydrogen and Gemini 4 Carbon and targets complex coding, analytical reasoning and enterprise workflows. It costs $3 per million input tokens ($0.15 cached) and $15 per million output tokens, has a 1M-token context and four thinking levels. Google reports #1 on the Artificial Analysis Coding Agent Index (58.6), #3 on the Intelligence Index (61 at xhigh), 70.9% on the Vals Index, 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0, at 32% lower cost per task than Carbon at high thinking. It is available on the paid Gemini API, AI Studio, Antigravity and the Gemini app; no weights.
Why it mattersNeon fills the price-performance middle tier that most production agents actually run on, directly against Claude Sonnet and GPT-6.2 Sol. The benchmark claims are Google's own, so independent evaluations over the coming days will matter.
Technical report for Clover-35B-A3B-Preview, a 35B mixture-of-experts diffusion language model with about 3B active parameters and a thinking mode; Apache-2.0 weights have been on Hugging Face since late October. It adapts Gated Delta Networks to block diffusion, combining recurrent memory of earlier context with bidirectional attention inside the current block. The authors report leading results on seven agent benchmarks, including 73.6 on SWE-bench Verified, and 680 tokens/s on a single GPU. A variant answers related questions in one pass, raising accuracy on an internal question-group set.
Why it mattersDiffusion LMs have mostly been research curiosities; this is an open-weight one claiming competitive agentic coding scores at high speed. The report finally documents how the hybrid recurrent-plus-block-diffusion design works, so others can study and reproduce it.
Induction Head Origins of SLAM-ASR Speech Recognition Models
SLAM-ASR maps speech encoder outputs into the embedding space of a frozen language model through a small trained projection. The authors show the model relies on induction heads, the same heads that copy repeated token sequences in text, to map speech embeddings to transcript tokens. They also find head pairs acting like a constrained differential attention that sharpens focus on the previously generated word, and show these heads can be used to improve transcription on noisy audio.
Why it mattersSpeech LLMs are widely used but poorly understood; this gives a concrete mechanism for why a frozen text LLM can transcribe at all. Knowing which heads do the work opens practical levers for robustness, as the noisy-audio result shows.
AssemblyAI-STT-3.7-Realtime-Agent: Speech-to-text built for voice agents
AssemblyAI released a streaming speech-to-text model built for voice agents, handling background speech, short replies, numbers and end-of-turn detection, at $0.45/hour. It is trained on tens of thousands of hours of real voice-agent and telephony calls and uses context from earlier turns, with no speaker enrollment. Final transcripts arrive 86 ms after the caller stops (vs 172 ms), WER is 4.53% on 12,460 English voice-agent calls (lowest of 26 streaming models tested), number errors fall 80% (11.28% to 2.27%), background leakage falls 47% and end-of-turn F1 rises from 83.60% to 85.52%. Most figures are AssemblyAI's own; Voice Arena ranks it #1 among streaming models for US English.
Why it mattersIn cascaded voice agents, ASR latency, missed short replies and misheard numbers are the failures users actually notice. A model tuned for exactly those cases, at the same price, raises the bar for the cascaded stack against end-to-end speech models.