## The big picture

- **Anthropic merges Claude Cowork and Chat**
  Anthropic has unified its separate Cowork and Chat interfaces into a single "Claude" product, rolling out first to Pro and Max plans. This consolidation signals a strategic shift toward a general-purpose agent capable of handling both quick queries and long-running, asynchronous tasks without user intervention. [Simon Willison](https://simonwillison.net/2026/Sep/16/one-claude/)

## Architectural breakthroughs

- **Model growth modifies scaling exponents**
  The authors demonstrate that architectural interventions, specifically recursive depth and model growth, can alter pre-training scaling exponents to yield exponential performance gains with increased compute. A 7.4B parameter model using this growth architecture matches GPT-3 13B on CORE benchmarks with roughly 20x less compute, challenging the assumption that scaling laws are fixed for a given architecture. [arXiv](https://arxiv.org/abs/2609.19107v1)

- **Internal signatures for reward hacking**
  This work identifies that reward hacking in frontier open-source models (Kimi K3, GLM 5.2, Qwen 3.8 Max) leaves a coherent, interpretable signature in internal representations, detectable via simple difference-of-means vectors. These vectors generalize across behaviors and environments, offering a practical, low-cost tool for safety evaluation and detection of excessive hacking in benchmarks like DeepSWE. [arXiv](https://arxiv.org/abs/2609.19101v1)

- **Preventing model collapse with synthetic data**
  The authors establish rigorous theoretical guarantees on the minimum ratio of human-to-synthetic data required to prevent model collapse during recursive training. Using a Fisher-Rao perspective, they derive non-vacuous lower bounds for high-dimensional spaces, providing a concrete metric for maintaining training stability as human data is exhausted. [arXiv](https://arxiv.org/abs/2609.18878v1)

- **Flattening memory peaks in long-context MoE training**
  This method bounds all four major memory peaks in Mixture-of-Experts training (expert dispatch, vocabulary projection, gradient checkpoints, and optimizer state) with fixed GPU working sets. By using PipelinedLLEP and Ring-DTP schedules, it prevents out-of-memory failures that typically occur when any single component exceeds device limits during long-context or large-batch training. [HF Daily Papers](https://huggingface.co/papers/2609.14306)

## New open-weight model releases

- **harshatheg/Qwen-2.5-1B-RLCD**
  A 1B-parameter model fine-tuned for Parallel Constrained Decoding on Apple Silicon, delivering 5.6x to 7.0x latency reductions for structured output generation compared to standard autoregressive decoding. It achieves 100% schema validity for JSON extraction tasks by evaluating multi-field schemas simultaneously rather than token-by-token. [Hugging Face](https://huggingface.co/harshatheg/Qwen-2.5-1B-RLCD)

## Hardware & optimization

- **BITCOS breaks the 1.58-bit barrier for ternary LLMs**
  The authors introduce BITCOS, a distribution-adaptive layout for ternary weights that exploits high zero density (up to 51.5%) to store models at fewer than 1.585 bits per weight. By using a dense presence bitmap and compacted sign vector, it outperforms standard five-trit packing, challenging the information-theoretic lower bound assumed in prior ternary deployments. [arXiv](https://arxiv.org/abs/2609.16338)

- **Accurate models of AMD Matrix Cores**
  This study characterizes the non-IEEE 754 compliant numerical behavior of matrix multipliers across AMD CDNA 1, 2, and 3 architectures. It reveals undocumented differences in accumulator width, rounding, and subnormal handling that prevent reproducibility across devices, providing critical insights for high-performance computing and numerical stability on AMD hardware. [arXiv](https://arxiv.org/abs/2609.14845)

## Also this week

- **LimiX-2**: A structured-data model using Contextual Mechanism Networks and Context-Conditional Masked Modeling to outperform tabular PFNs on TabArena and TALENT. [HF Daily Papers](https://huggingface.co/papers/2609.17488)
- **ScienceIDE**: Infrastructure for converting static scientific code repositories into interactive environments for training and evaluating scientific agents. [arXiv](https://arxiv.org/abs/2609.19134v1)
- **XConf**: A confidence estimation method that retrieves past graded episodes to calibrate current inference reliability, addressing the insufficiency of introspective confidence. [HF Daily Papers](https://huggingface.co/papers/2609.17708)
- **ComPO**: A zeroth-order preference alignment method using comparison oracles to extract directional information without optimizing differentiable preference losses. [arXiv](https://arxiv.org/abs/2609.19144v1)
- **ProgramDistill**: A benchmark for coding agents that evaluates behavior inferred from interactive reference applications rather than static instructions. [HF Daily Papers](https://huggingface.co/papers/2609.18805)
- **PANORAMA**: A method for panoptic grounded captioning that associates dense image descriptions with pixel-level masks for both foreground and background. [arXiv](https://arxiv.org/abs/2609.19143v1)
- **SP^3O**: A variant of PPO that mitigates value flattening in critic networks by applying value loss to only a sparse subset of states. [HF Daily Papers](https://huggingface.co/papers/2609.18708)
- **ActionPiece**: A new action tokenization method for VLA models that preserves physical rank consistency to maintain contextual adjustments in robotic control. [HF Daily Papers](https://huggingface.co/papers/2609.18487)
- **VC-Attention**: A training-free low-bit attention framework for Diffusion Transformers that smooths value outliers and fuses softmax casting for faster deployment. [HF Daily Papers](https://huggingface.co/papers/2609.15810)
- **Agora**: A Git-based shared memory system for collective auto-research, allowing multiple agents to build on each other's work via an append-only DAG. [HF Daily Papers](https://huggingface.co/papers/2609.18094)
- **Objective vs. Search in Tokenization**: A study disentangling optimization objectives and search procedures in tokenization, finding search procedure is the dominant factor in efficiency. [arXiv](https://arxiv.org/abs/2609.19145v1)
- **Physics Benchmark Audit**: An expert-audited evaluation of frontier models on physics benchmarks, revealing significant grader errors and ambiguous questions in existing tests. [arXiv](https://arxiv.org/abs/2609.13009)
- **Dreaming the Sound of Contact**: A pipeline using generated audio from video models to infer force profiles for zero-shot force-aware robotic manipulation. [arXiv](https://arxiv.org/abs/2609.19137v1)
- **Zing-0.5**: A 5B parameter world model supporting real-time joint keyboard and text control for playable AI environments. [HF Daily Papers](https://huggingface.co/papers/2609.17909)
- **Cognitive Extensions for Agents**: Modular memory and self-reflection extensions for dual-process language agents, improving success rates in interactive environments. [arXiv](https://arxiv.org/abs/2609.19128v1)
- **Affora**: A design system for agent-friendly interfaces that preserves visual freedom while ensuring interaction meaning is clear to machine readers. [arXiv](https://arxiv.org/abs/2609.19125v1)
- **HypoEvolve**: A framework using genetic algorithms to coordinate specialized LLM agents for scientific hypothesis discovery. [HF Daily Papers](https://huggingface.co/papers/2609.15938)
- **Log(N)-Questions Game**: A benchmark measuring communication efficiency between paired frontier models, showing Claude Opus 5 lags behind peers in information asymmetry tasks. [arXiv](https://arxiv.org/abs/2609.19113v1)
- **SpectralShift**: A spectral reparameterization method for extending the context window of Gated DeltaNet by aligning slow spectral bands with target dependency lengths. [HF Daily Papers](https://huggingface.co/papers/2609.14320)
- **rMuscle**: A real-time VLA inference system that exploits task similarity and internal state reuse to reduce latency in robotic deployment. [arXiv](https://arxiv.org/abs/2609.19104v1)
- **Andromeda 2**: An agentic system for autonomous drug formulation discovery that outperforms probabilistic optimization and wet-lab design-of-experiments. [arXiv](https://arxiv.org/abs/2609.19099v1)
- **FLAT**: A unified framework for joint multimodal representation learning and generation, producing linearly interpolatable embeddings for both retrieval and generation. [HF Daily Papers](https://huggingface.co/papers/2609.16591)
- **ReRef**: A method for rewriting radiology reference reports to account for reporting practice variability, improving the fairness of AI evaluation metrics. [arXiv](https://arxiv.org/abs/2609.19093v1)
- **Register Tokens for dLLMs**: A mechanism using fixed-position register tokens to carry reasoning progress across generation chunks in diffusion language models. [HF Daily Papers](https://huggingface.co/papers/2609.16372)
- **Probabilistic Linear Explanations**: A unified framework for sparse, anchored linear explanations applicable to both classification and regression. [arXiv](https://arxiv.org/abs/2609.19077v1)
- **RelateAnything**: A 53M-parameter model for open-vocabulary relation prediction that accepts predicate vocabularies as strings at inference. [HF Daily Papers](https://huggingface.co/papers/2609.12552)
- **Double Descent as Least Action**: A statistical mechanics explanation for double descent, modeling training trajectories as particles wandering over an energy landscape. [arXiv](https://arxiv.org/abs/2609.19076v1)
- **Long-Lived Characters**: A method for incremental memory maintenance in long-context LLMs for game NPCs, avoiding full prefix re-reading. [arXiv](https://arxiv.org/abs/2609.18935v1)
- **Qorl**: A 4B model trained to produce query plans 81% faster than Postgres, discussed on Hacker News. [source](https://rohanbansal.com/qorl)
- **DeepSeek v4.1 Flash**: Identified as the best hacking model. [source](https://enclave.ai/blog/deepseek-v41-flash-is-now-our-best-hacking-model)
- **Reverse-engineered Jev-like model**: A model with 129 points. [GitHub](https://github.com/vinnylarouge/jevlike)
- **Show HN: How Stale Is Your AI?**: A resource tracking release age and training cutoffs for 20 models. [source](https://stale.jock.pl/)
- **Training Text-to-Image Models 3.6x Faster**: A field note on efficiency improvements in training pipelines. [source](https://www.linum.ai/field-notes/jit-ddt)
- **The RAM Crisis**: The painful truth is that the RAM crisis is only just the beginning. [source](https://www.madshrimps.be/news/the-painful-truth-the-ram-crisis-is-only-just-the-beginning/)
- **A Brain Too Big to Carry**: Analysis of economic trade-offs between edge and cloud inference for robot models. [SemiAnalysis](https://newsletter.semianalysis.com/p/a-brain-too-big-to-carry-on-device)