You ghost your own socials by Wednesday. SureThing learns your voice and ships native posts to LinkedIn, X, Instagram, and TikTok, without you writing a thing.

Post in Your Voice

Mozer et al. [Google DeepMind, University of Texas]

♥ 989 LLM Memory

Finding efficient ways to enhance an AI's internal memory without retraining massive models from scratch could unlock far more reliable reasoning and context awareness.

AI models process text in a one-way forward pass from shallow to deep layers. Since shallower processing stages cannot access the contextual insights calculated deeper in the network, AI models struggle to update their internal understanding, which leads to reasoning failures and context errors over time.

To address this limitation, researchers introduced a technique called recirculation, which routes a tiny fraction of deep-layer activation back into shallower layers as the model moves step-by-step through text.

By leaking these refined context signals downward, the network acts like a dynamical system capable of maintaining persistent belief states over time. Across multiple diverse model families, this simple feedback loop significantly improves text predictability out of the box without modifying any of the underlying model weights.

The approach proves especially effective for extended generative reasoning, dramatically boosting accuracy on multi-step math problems and complex instruction-following tasks.

On top of this, the team also developed an adaptive variant that uses small secondary networks to dynamically control how signals blend between layers. By tuning only these tiny mixing parameters while keeping the primary language model entirely frozen, the architecture achieved remarkable performance gains, including substantial error reductions on mathematical reasoning benchmarks and widespread improvements in language fluency.

Liu et al. [University of Washington, Stanford University, Northeastern University, Carnegie Mellon University, Massachusetts Institute of Technology, National University of Singapore, Seoul National University, Stevens Institute of Technology, University of Chicago]

♥ 657 LLM RL bycloud’s pick

AI training methods use fixed problem sets which are quickly mastered by AI during the training stage and this creates a hard ceiling on continuous self-improvement. To solve this, researchers introduced SPADE, a self-play framework where a single AI model continuously teaches itself by alternating between two roles: an environment designer and a reasoning agent.

In this, the modesl operates as the designer and it writes a complete, multi-step training environment as runnable Python code, which contains everything from complex logic games to stateful digital tool-use tasks.

The designer attaches a privileged hint to each task and measures how much easier the hint makes the problem. If the agent easily succeeds without the hint, the task is too simple; if it fails even with the hint, the task is impossible.

The highest rewards are given to system which are only solvable after a hint is passed, this ensures that the training material automatically targets the frontier of the agent's current abilities.

The designer uses a diverse corpus of real-world scientific documents and maintains a memory of past successful tasks. As the reasoning agent grew smarter, the designer naturally evolved its tasks from simple, static puzzles into multi-turn challenges that forced the model to actively gather evidence before drawing conclusions.

Dong et al. [Stanford University, Peking University]

♥ 507 World Models

RL allows AI models to master real world tasks but, training AI agents quickly and efficiently is hard. World models can perform some intricate actions in real world environments, but traditional approaches force agents to learn directly inside these virtual simulations. This often backfires because small prediction errors in the simulated environment compound over time, and this causes the agent's strategy to break down during long or visually complex tasks.

To solve this, researchers introduced Q-Learning with World Models. This framework changes how world models are used. Rather than training the agent inside an imperfect simulation, QWM keeps the core policy and value functions grounded in real interactions.

The world model is deployed purely at decision time to let the agent temporarily "imagine" short sequences of future actions before committing to a move. By conducting a targeted search over candidate actions and evaluating their predicted future outcomes, the agent selects the highest-value action while completely avoiding the buildup of simulation errors.

This works both while the agent gathers new training data and during final evaluation, this ensures higher-quality experiences are collected throughout the entire learning process.

QWM outperformed existing state-of-the-art methods in both learning efficiency and overall success rates across robotic benchmarks which contain both low-dimensional state data and complex camera video inputs.

Cheng et al. [STAR, Weizmann Institute of Science, MIT-IBM Watson AI Lab, University of Cambridge, EleutherAI]

♥ 855 LLM Behaviors

AI models problem solving abilities should remain unchanged, regardless of what language the user speaks. We know that AI models access stored facts unevenly across languages but it is unclear whether a model’s underlying reasoning, planning, and strategic skills also change simply by switching its communication language.

To isolate this issue, researchers created a new evaluation method which uses competitive text-based games. By pitting two identical instances of the same AI model against each other, one operating in English and the other in a language like German, Arabic, or Malay, the team could observe pure skill differences.

Because the board states, rules, numbers, and allowed moves remained identical, any gap in win rates exposed how language alone affects decision-making.

These experiments showed us that the same models have different strengths depending on its language interface. In testing, English consistently brought out the strongest strategic play. Whereas, interfaces using non-Latin scripts, such as Arabic and Hebrew, caused the model to struggle with spatial reasoning and fail to recall mathematical strategies that it easily executed in English.

Researchers found that when a model was instructed to perform its intermediate reasoning steps in a stronger language before answering in a weaker interface language, it recovered most of its lost performance.