Reasoning

BPO turns policy mirror descent into a critic-free trajectory objective for terminal rewards, avoiding intermediate value estimates. It is relevant to builders training reasoning models with verifiable rewards.

Tiny Aya L2-Thinker shows multilingual reasoning can transfer through data mixing, suggesting builders should evaluate whether agents reason in the user's language, not only answer in it.

RLMs keep large inputs in a programmable environment, letting a model inspect them with code and delegate subsets to submodels instead of attending to every token at once.

GPT-6 Astra combines reasoning, computer use, and writing and design judgment, potentially widening the range of knowledge-work tasks one agent can handle.

Molecular benchmark accuracy can reflect retrieval of published values rather than prediction. Contamination audits should test digit-level recall and repeat runs across reasoning settings.

A compact classifier distilled from LLM rationales handled product-pair decisions without inference-time LLM calls. Category adapters improved accuracy further while retaining large speed and cost gains.

MiniMax argues that long-running, multimodal agents need enough context for repeated tool responses, while sparse attention makes that context more tractable.

Readable chain-of-thought is not a reliable map of which steps caused a correct answer. Treat step-level LLM critiques as partial signals, especially on already-correct traces.

Controlled pre-training experiments suggest varied reformulations can teach facts more efficiently than repeating documents under the same token budget, though paraphrasing gains depend on batch size.

On-policy distillation recovered most full-data gains from one query; diverse rollouts mattered more than dataset size, while slow student alignment remained the bottleneck.

Legora says GPT-6 Astra reviewed 41 documents within minutes, caught every planted error, and improved its workflow result by nearly 40%, though the underlying measure is unspecified.

OpenAI introduces GPT-6 Astra with claimed advances in computer use, coding, cybersecurity, and science. The supplied material gives no benchmarks or implementation details to assess those gains.

A coding-specialized model paired post-training with an iterative generate-evaluate-refine loop to exceed the top IOI 2026 human score. The reusable idea is feedback-driven test-time search.

Cliff trains on the location of a rollout’s first reasoning error, rewarding the valid prefix and penalizing the suffix. It offers finer supervision without a specialized reward model.

Gemini 3.8 Flash brings multimodal input, tool calling, web search, and default reasoning to coding agents through Vercel. Its temporary 50% discount runs through December 31.

DARTS targets representation drift in merged decoder LLMs with entropy-weighted, position-aware correction, adding 0.1% parameters in the reported Llama-2-7B tests.

CritICL turns recurring failures from smaller models into critique examples for stronger models, aiming to improve inference with fewer generations and lower token cost than repeated sampling.

TTPO adapts a model at test time without labels by treating agreement and disagreement differently, reducing the damage from wrong majority-vote pseudo-labels.

Feeding a model partial reasoning from a weaker model during RLVR preserved exploration and improved large-k coverage. This suggests diversity can come from cross-model perturbation, not only regularization.

Three ways to combine RLVR domain experts perform similarly on average but diverge by task. Choose Merge for cheap reuse, Mix RL for training from pooled data, and MOPD for preserving expert gains.

Ling 3.0 Flash Fin adds a finance-focused reasoning and tool-calling option to AI Gateway, with separate model IDs for automatic billing or a hard stop after the free period.

Prefix Sliding drops stale reasoning tokens while retaining instructions and recent work, capping memory during long inference and reportedly making existing models 3× faster without training.

AI4AI-Bench tests whether agents can rewrite training algorithms, not merely tune runs. The best system scored 0.250, with most submissions never changing how the model learns.

A model learned to choose its own reasoning budget, cutting average output tokens 41% with a small MATH500 accuracy drop. Agent builders could replace fixed effort settings with task-level routing.

Reactor frames real-time video as a programmable session rather than a generated file, enabling interactive worlds and live editing but exposing hard state, latency, and evaluation problems.

GLM 5.3 is available through Vercel AI Gateway for coding agents, retaining a 1M-token context window while claiming better long-horizon engineering with fewer output tokens.

YOPO combines reasoning steering and answer abstention in one pass of a frozen model. Reconstructing the clean residual avoids interference and beats the paper's two-pass reference at three model sizes.

LittleLearner offers a controlled model and corpus for studying knowledge acquisition without unknown prior exposure. Its initial results separate better use of known material from new capability.

SAEVerbalizer turns sparse-autoencoder decoder directions into natural-language feature explanations. It may reduce behavior-sampling work when inspecting model representations, but evidence is experimental.

Mimir v1 is an open 1B-parameter reasoning model trained with permissible post-training data. It is relevant for compact or Danish deployments, though the abstract supplies no benchmark scores.

Targeted self-distillation taught a coding model to finish within its turn budget without reducing test pass rate. The practical levers were step-specific hints and selective token learning.

Auto Scientist aims to automate model-training choices across data, alignment, and architecture. The builder-relevant claim is broader recipe search, though frontier training remains compute-heavy and safety stays unresolved.

Grok 4.6 targets long-running coding and knowledge-work agents, with more self-testing and stronger visual first passes reported by Cursor. API pricing starts at $2 input and $6 output per million tokens.

CreativeInstruct adds learned control spans that recover base-model-like diversity after post-training, with reported gains in human creativity ratings and downstream RL training.

RP-OPSD targets cross-lingual distillation at tokens that steer reasoning rather than surface wording, a useful training pattern for multilingual reasoning models.

RRC converts a generative reward model’s relative rankings into RL rewards, aligning the training signal with what the evaluator naturally produces instead of forcing direct scalar scores.

OctoLong builds dependency-rich, cross-repository training contexts by following code references. The results suggest long-context coding models benefit from linked program structure, not just larger text windows.

Skill²-Bench measures whether models can switch reasoning skills across dependent steps. Its results expose a failure mode relevant to agents whose plans mix coding, retrieval, calculation and tool use.

OPD-V selects self-distillation tokens by measuring whether visual input influences generation, addressing text dominance that can blunt multimodal reasoning training.

ParVL reuses shared vision and language backbones across parallel branches, letting multimodal systems shift extra compute between modalities without adding separate backbone parameters.