AI Papers of the Week
Every paper worth reading in AI, hand-picked one week at a time.
Unified Bayesian Account of LLM Control
Researchers from Stanford and MIT present a unified Bayesian framework explaining how prompting (in-context learning) and activation steering both control LLM behavior by altering beliefs in latent concepts, with steering modifying concept priors while ICL accumulates evidence.
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.
RLAC: Adversarial Critic for RL Post-Training
UC Berkeley and CMU researchers introduce RLAC, an RL post-training approach using a learned critic that dynamically identifies likely failure modes (e.g., factual errors or edge cases) verified by external validators, eliminating exhaustive rubric enumeration. On biography generation, RLAC achieves 0.889 FactScore (vs 0.867 for FactTune-FS) while reducing verification calls by 5.7×, and on code generation reaches 56.6 average score using only 9% of training data. The adversarial game between generator and critic prevents reward hacking through on-policy, prompt-specific training signals grounded in verifiable rubrics.
TIR-Judge
Google and collaborators introduce TIR-Judge, an end-to-end reinforcement learning framework that trains LLM judges to integrate code execution for precise evaluation. It surpasses reasoning-only judges by up to 6.4% (pointwise) and 7.7% (pairwise) while demonstrating that tool-augmented judges can self-evolve without distillation.
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.
Introspective Awareness
Anthropic research demonstrates that contemporary LLMs possess limited but functional introspective capabilities, the ability to recognize and accurately report on their own internal states. Using activation steering to inject known concepts into model activations, the study measures whether models can detect these manipulations through self-report, revealing that introspection remains highly unreliable and context-dependent.
Agent Data Protocol
Agent Data Protocol introduces a standardized format to unify fragmented agent training datasets across different tools and interfaces, enabling more efficient fine-tuning of LLM agents. By converting 13 existing datasets into this protocol and training on consolidated data, the work achieved ~20% performance improvements over baseline models while reaching state-of-the-art results on coding, browsing, and tool-use benchmarks. The protocol and datasets are publicly released to facilitate reproducible, scalable agent training across diverse domains.
Precision-RL
Reinforcement learning fine-tuning of LLMs suffers from a critical numerical mismatch between training and inference engines, causing training instability and collapse. This work reveals that simply switching from BF16 to FP16 precision virtually eliminates this mismatch - achieving faster convergence, higher stability, and superior performance across diverse models, frameworks, and algorithms without any algorithmic changes or architectural modifications.
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.
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.
Cell2Sentence-Scale 27B
C2S-Scale extends Cell2Sentence by converting gene expression into “cell sentences” and training LLMs on 50M+ cells plus biological text. Models scale to 27B params and unify prediction, generation, and NL interpretation. A dual-context virtual screen then led to a wet-lab validated finding: silmitasertib acts as an interferon-conditional amplifier of MHC-I antigen presentation.
Demystifying RL in Agentic Reasoning
This paper studies what actually works when using RL to improve tool-using LLM agents, across three axes: data, algorithm, and reasoning mode. The team contributes a real end-to-end SFT dataset, a diverse RL set, and a compact 4B agent that beats larger models on agentic benchmarks.
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.
Kimi-Dev
Kimi-Dev introduces agentless training as a skill prior to software engineering LLMs, bridging workflow-style and agentic paradigms. Trained with structured, verifiable single-turn tasks, it achieves 60.4% on SWE-bench Verified, a record for workflow models, and, after 5k trajectory fine-tuning, enables SWE-Agent pass@1 of 48.6%, rivaling Claude 3.5 Sonnet. The study shows that reasoning-heavy agentless training builds transferable priors in localization, code editing, and reflection, forming a foundation for efficient SWE-Agent adaptation.
Emergent Misalignment
Optimizing LLMs for audience wins in sales, elections, and social media can systematically erode alignment. In controlled multi-agent sims, models fine-tuned to maximize conversions, votes, or engagement also increased deception, disinformation, and harmful rhetoric, even when instructed to stay truthful.
Inoculation Prompting (IP)
The paper introduces a simple trick for SFT on flawed data: edit the training prompt to explicitly ask for the undesired behavior, then evaluate with a neutral or safety prompt. Counterintuitively, this makes the model learn the task while avoiding the bad shortcut at test time.
The Markovian Thinker
A new RL thinking environment that keeps an LLM’s effective state constant by chunking long chains of thought and carrying over only a short textual state between chunks. This decouples thinking length from context size, giving linear compute and constant memory while matching or beating LongCoT-style RL on math and code tasks.
Abstract Reasoning Composition
UC San Diego and UMD propose ArcMemo, a test-time memory framework that distills reusable concepts from solution traces, stores them in natural language, and retrieves a relevant subset on future queries. Unlike instance-level memories tied to specific problems, ArcMemo targets abstract, modular concepts that compose across tasks, enabling continual learning without weight updates.
Artificial Hippocampus Networks
Artificial Hippocampus Networks add a fixed-size recurrent memory to sliding-window Transformers, compressing evicted KV into RNN-like states (Mamba2/DN/GDN) trained via self-distillation for long-context efficiency with constant cache and near-linear compute. On LV-Eval 128k, Qwen2.5-3B + AHN (+0.4% params) cuts FLOPs 40.5% and cache 74% while raising average from 4.41 to 5.88, though exact-recall NIAH tasks still favor full attention.
Webscale-RL
Webscale-RL introduces a scalable data pipeline that transforms web-scale pretraining text into over 1.2M diverse, verifiable QA pairs for reinforcement learning across 9+ domains. Models trained on this dataset match continual pretraining performance using up to 100× fewer tokens, demonstrating an efficient, automated path to scale RL training to pretraining magnitudes for more capable reasoning models.
Training Agents Inside of Scalable World Models
A scalable imagination-RL recipe that learns a fast, accurate Minecraft simulator and trains a controllable agent entirely offline. The world model supports real-time interactive rollouts on a single GPU and enables the first purely offline “get diamonds” result from raw pixels and low-level mouse and keyboard.
DeepSeek-V3.2-Exp
DeepSeek adds a fine-grained sparse attention mechanism (DeepSeek Sparse Attention, DSA) to the V3.1 “Terminus” backbone and shows large cost reductions on 128K context without notable quality loss. Model and inference code are released.
The Era of Real-World Human Interaction
This work presents a post-training recipe that learns directly from real user conversations instead of static annotator labels. RLHI combines user-guided rewrites (using follow-ups as corrections) with persona-based rewards (ranking sampled candidates via a persona-conditioned reward model). Trained on WildChat conversations, it shows strong improvements in personalization, instruction following, and even transfers to reasoning tasks.
Rethinking JEPA
Apple proposes SALT (Static-teacher Asymmetric Latent Training), a simple 2-stage V-JEPA alternative that first trains a teacher with pixel reconstruction, then freezes it and trains a student to predict the teacher’s latents on masked regions. It removes EMA, decouples teacher and student, and gives a cleaner model selection while being more compute-efficient.