AI Papers of the Week
Every paper worth reading in AI, hand-picked one week at a time.
MACI
This paper argues that LLMs are not fundamentally limited as pattern matchers - the real bottleneck is the lack of a System-2 coordination layer. The authors propose MACI, an architecture implementing three mechanisms: baiting (behavior-modulated debate), filtering (Socratic judging), and persistence (transactional memory) to enable goal-directed reasoning on top of LLM substrates.
Test-Time Training for Long-Context LLMs
This paper shows that long-context LLMs can access millions of tokens but often fail to meaningfully use that information. The authors propose query-only test-time training (qTTT), which adapts models during inference through targeted gradient updates rather than generating more thinking tokens.
Epistemia
This paper argues that LLMs are not epistemic agents but stochastic pattern-completion systems. By mapping human and artificial epistemic pipelines, the authors identify seven fundamental fault lines where human and machine judgment diverge, despite producing superficially similar outputs.
Empirical Study of Agent Developer Practices
This paper presents the first large-scale empirical study of LLM-based agent frameworks, analyzing 11,910 developer discussions across ten popular frameworks. The research identifies practical challenges developers face and evaluates how well current frameworks meet their needs.
Sophia
Sophia introduces System 3, a meta-layer beyond traditional dual-process theory that enables LLM agents to maintain persistent identity and align short-term actions with long-term goals. The framework achieves 80% reduction in reasoning steps for recurring operations and 40% performance improvement on high-complexity tasks.
SonicMoE
SonicMoE addresses performance bottlenecks in Mixture of Experts models through IO-aware and tile-aware optimizations. The approach achieves 1.86x compute throughput improvement on Hopper GPUs, reduces activation memory by 45%, and enables training 213 billion tokens per day on 64 H100 GPUs for a 7B model.
DeepCode
DeepCode is a fully autonomous framework for synthesizing complete codebases from scientific papers despite LLM context limitations. It treats repository synthesis as a channel optimization problem, achieving state-of-the-art on PaperBench and outperforming commercial tools like Cursor and Claude Code.
CLaRa
CLaRa introduces a unified framework for retrieval-augmented generation that performs embedding-based compression and joint optimization in a shared continuous space. The approach addresses key RAG limitations around long contexts and disjoint retrieval-generation optimization.
DeepSeek-V3.2
DeepSeek releases V3.2, an open model that matches GPT-5 on reasoning benchmarks while introducing significant architectural and training innovations. The high-compute variant, DeepSeek-V3.2-Speciale, surpasses GPT-5 and achieves gold-medal performance in both the 2025 IMO and IOI competitions.
LatentMAS
LatentMAS introduces a framework enabling language model agents to collaborate directly within a continuous latent space rather than relying on text-based communication. By using last-layer hidden embeddings and a shared latent working memory, agents preserve and transfer internal representations without information loss from text serialization.
OLMo 3
Allen Institute for AI introduces OLMo 3, a fully open language model family that releases the complete “model flow”: every training stage, checkpoint, dataset, and dependency, enabling researchers to intervene at any development point. The release includes four specialized variants (Base, Think, Instruct, RL Zero) at 7B and 32B scales.
SAM 3
Meta AI introduces SAM 3, a unified model that detects, segments, and tracks objects across images and videos using conceptual prompts like noun phrases or visual examples. This extends the Segment Anything capability to concept-based segmentation through Promptable Concept Segmentation (PCS).
Nested Learning Framework
Google Research introduces Nested Learning (NL), a paradigm representing models as nested optimization problems where each component has its own context flow, revealing that deep learning methods compress context and explaining how in-context learning emerges. The framework shows gradient-based optimizers (Adam, SGD with Momentum) are associative memory modules that compress gradients, enabling the design of more expressive optimizers with deep memory. The HOPE architecture, combining self-modifying sequence models with continuum memory systems, achieves strong results on language modeling (15.11 WikiText perplexity at 1.3B parameters), outperforming Transformers and modern recurrent models.
Context Engineering 2.0
Researchers from SJTU, SII, and GAIR trace the 20+ year evolution of context engineering, reframing it as a fundamental challenge in human-machine communication spanning from primitive computing (Era 1.0) to today’s intelligent agents (Era 2.0) and beyond. It defines context engineering as systematic entropy reduction where humans preprocess high-entropy contexts into low-entropy machine-understandable representations. This gap narrows as machine intelligence increases.
Enhancing Long-Term Memory in LLMs
Researchers from the University of Alberta and UMass Amherst introduce BEAM, a new benchmark for evaluating long-term memory in LLMs with conversations up to 10M tokens, and LIGHT, a framework that enhances memory performance through three complementary systems.
AgentFold
AgentFold introduces proactive context management for long-horizon web agents, addressing context saturation through dynamic “folding” operations that balance detail preservation with efficient compression. The 30B parameter model outperforms dramatically larger competitors while achieving state-of-the-art results on web browsing benchmarks.
Kimi Linear
Kimi Linear introduces a hybrid linear attention architecture combining Kimi Delta Attention (KDA) with periodic full attention layers at a 3:1 ratio, achieving superior performance over full attention while reducing KV cache by 75% and delivering 6× faster decoding at 1M context. KDA extends Gated DeltaNet with fine-grained channel-wise gating and specialized Diagonal-Plus-Low-Rank matrices, enabling more effective RNN memory management while maintaining hardware efficiency through optimized chunkwise algorithms that substantially reduce computation versus general DPLR formulations.
DeepSeek-OCR
DeepSeek-OCR explores compressing long text contexts into visual representations using a novel vision encoder architecture (DeepEncoder) that achieves 10-20x compression ratios while maintaining high OCR accuracy.
Continual Learning via Sparse Memory Finetuning
Meta AI researchers address catastrophic forgetting in language models through sparse memory finetuning, updating only memory slots most activated by new knowledge while achieving 89% less performance degradation than standard finetuning.
Bayesian Influence Functions for Hessian-Free Data Attribution
Classical influence functions struggle with deep neural networks due to non-invertible Hessians and high-dimensional parameter spaces. This work introduces the local Bayesian influence function (BIF), which replaces Hessian inversion with loss landscape statistics estimated via stochastic-gradient MCMC sampling.
Ring-1T
Ring-1T is the first open-source thinking model with 1 trillion parameters (~50B active per token), achieving breakthrough results through three innovations for trillion-scale RL training.
Elastic-Cache
A training-free, architecture-agnostic way to make diffusion LLM decoding fast by updating KV caches only when and where it matters. Instead of recomputing QKV for all tokens at every denoising step, Elastic-Cache watches attention drift on the most-attended tokens and refreshes only deeper layers while reusing shallow and off-window caches. Results: large speedups with minimal or no accuracy loss across math, code, and multimodal tasks.
LLMs Can Get “Brain Rot”!
The authors test a clear hypothesis: continual pretraining on trivial, highly engaging web text degrades LLM cognition in ways that persist even after mitigation. They build controlled Twitter datasets to isolate data quality from scale and training ops, then measure effects on reasoning, long-context, safety, and personality.
Agentic Context Engineering (ACE)
Presents a modular context-engineering framework that grows and refines an LLM’s working context like a playbook, not a terse prompt. ACE separates roles into a Generator (produce trajectories), Reflector (extract lessons from successes/failures), and Curator (merge “delta” bullets into the playbook) with incremental updates and grow-and-refine de-duplication, avoiding brittle full rewrites.