The week’s most important papers on large language models: reasoning, agents, evaluation and alignment — with real results, not position pieces.
This first issue has 11 items. Reasoning and training methods lead: one paper shows that fixed opening tokens recover much of RL's gain in base models, and another distills power-sampled answers into single generations. TasteVal, a benchmark on which Opus 5.5 beats human experts at designing experiments, is also worth reading. I left out MatrixFormer, which is a matrix-completion model rather than a language model.
Fixing a starting token cue makes base models competitive with their RL-trained counterparts on math and coding. The cue ".\n\nOkay" lifts Olmo-3-7B on MATH-500 from 42% to 78%. Causal data interventions trace the effect to training data, and a safety case study shows different cues elicit different refusal and compliance behavior.
Looped language models get cheaper when recurrent states sit near fixed points, which enables truncated backpropagation, terminal KV sharing, a distilled student that prefills up to 1.79x faster, and RL updates 2x faster. A learned depth prior and orthogonal injection lower perplexity at every scale from 100M to 1.6B. At 1.6B the learned prior with a 3x smaller KV cache matches fixed-depth training with the full cache.
MemPilot trains a multi-step LLM policy with RL to choose between retrieving from query-agnostic memory and delegating query-specific curation of raw multimodal history to heterogeneous LLMs and VLMs. It uses objective-wise advantage decoupling and prefix-based marginal utility estimation. On five multimodal agent-memory benchmarks, preference sweeps yield broader performance–cost–latency frontiers than existing trade-off-aware baselines.
Streams that already watch this field. Follow one as it is — it costs nothing extra.
Model launches, research, partnerships and funding from OpenAI, Google DeepMind, Anthropic, Meta AI, Mistral and NVIDIA.
New open-source frameworks and tools for building LLM agents, as they appear on GitHub and Hugging Face.
New open-weights language and multimodal models from the major labs, as they land on Hugging Face — with size, license and reported results.