#Reasoning

AILiquid AI shared a Hugging Face guide to multi-harness reinforcement learning for coding agents, in which a proxy records the token ids and logprobs vLLM samples so training works without changing the harness. Per the quoted post, LFM2.5-2.6B rose from 42% to 54% after training across four harnesses at once, and imitation fine-tuning on 3,189 rollouts from Qwen3.8-27B plateaued at 47.5%, below both RL runs. The proxy, trainer, tasks, SFT data, training code and seven trained models are described as open.

AIHugging Face and collaborators published a guide to multi-harness RL that trains models through a capture proxy without changing the agent harness. The proxy records the token ids and logprobs vLLM samples, and the source reports LFM2.5-2.6B rising from 42% to 54% after training across four harnesses. Fine-tuning on 3,189 successful rollouts from Qwen3.8-27B plateaued at 47.5%, below both RL runs, and the capture proxy, trainer, tasks, SFT data, training code, and seven trained models are released openly.

Why it matters: The source gives a concrete method for training models across several agent harnesses, with measured gains and a note that imitation learning underperformed RL.

AIThomas Wolf shares Kevin Buzzard's reflections on whether mathematics is about human understanding and what exponential AI growth means for an infinite field. He quotes Buzzard's view that machines may eventually reach a natural boundary where further progress is not worth the resources, and that humans would then take over from there.

AIThore Graepel, a core member of the AlphaGo team, argues that current large language models do not truly reason, despite chain-of-thought gains in math and coding. He says they lack an explicit, inspectable epistemic state, keep knowledge and reasoning intertwined in their weights, and often produce post-hoc explanations. He proposes systems that maintain an auditable epistemic state and evaluate each step by how much it resolves uncertainty.

AICool eval. Simply ask an LLM “Land or Water?” and give it a latitude and longitude coordinate as text. Ask 16,200 times, plot as image. The models know. From compressing the internet.

AIFrançois Chollet argues the key difference between base LLMs and modern LRMs is a shift from transductive answer prediction to inductive prediction of the program or reasoning chain behind an answer. He says this enables test-time induction and substantial fluid intelligence in LRMs, which he claims base LLMs largely lack. He cites ARC 1 results: base LLMs remain around 10-15%, while LRMs of the same size or smaller saturated the benchmark in 2025.

AIFireworks AI's Ember-1 produced a near-identical composition to Kimi K3 on the same JavaScript painting prompt, according to Sophia Yang. Ember-1 used 9x fewer thinking tokens and ran 4x faster at half the cost.

AIOdyssey introduces PROWL-2, a system in which agents and their world model improve through recursive learning. The post says agents expose errors in imagination, and repairing those errors enables further learning. Odyssey frames this open-ended learning as a critical step toward superintelligence.

AIOdyssey introduces PROWL-2, a recursive loop that delivers up to 91% relative gains over the StarCraft world-model baseline. The post also reports stronger robot coordination in simulation.

AIApodex has released Apodex 1.1 Mini, a free reasoning-first model designed for complex, long-horizon research and forecasting tasks. According to the source, it works directly with files, data, code, and tools to produce verifiable results.

AIZyphra published a paper and blog explaining how local mixing encodes relative position in global NoPE attention models. The work, titled "How Local Mixing Encodes Relative Position in Global NoPE Attention," is available on arXiv and on Zyphra's website.

AIZyphra says its results clarify how state-of-the-art NoPE models encode position and which inductive biases support generalization. It adds that global NoPE could enable models to extrapolate to contexts longer than those seen in training, potentially indefinitely.

AIModels using Kimi Delta Attention show a recency bias, as the mechanism controlling how much earlier information is kept shapes memory. As new information mixes into that memory, nearby words share more information, which gives the model clues about how far apart words are.

AIZyphra derives a mathematical theory showing how local layers produce recency bias that propagates through the residual stream, norms, and MLPs into NoPE attention logits. The authors validate the theory on both randomly initialized and trained networks.

AIHybrid NoPE models combine sliding window attention or recurrent layers, which focus on nearby words, with global attention layers that use no positional encoding (NoPE). The post notes that NoPE layers receive no positional information yet can still learn long-range dependencies, and raises the question of how this works.

AIZyphra says that mixing information within overlapping neighboring windows makes nearby words carry similar information inside a model. This creates a built-in clue about how far apart words are, even before training begins, which the post calls recency bias.

AIZyphra Research explains how language models track word order without explicitly encoding position in attention. The post says local memory layers that read nearby words help global attention layers preserve sequence information. The source is a short teaser thread, so no further technical details are given.

AIThe SYNTH paper, titled It's All Training, presents a fully synthetic single-stage pipeline for training workable reasoning models with high data efficiency. The authors argue this approach does not separate training into pretraining, mid-training, or post-training stages. The image shows the paper's abstract, which describes a pipeline built from a 58,000-article Wikipedia-based synthetic corpus and models named Baguettotron-600M and Baguettotron-MoE.

AIPareto 26.10 Preview is a multimodal composite model built for research, coding, and agentic workflows. It is described as delivering frontier-level performance across a broad range of general-purpose tasks, though the source excerpt is a preview and provides no benchmark scores, parameter counts, pricing, or availability details.

AIAmazon Science describes a graph-centric approach in which a network digital twin graph and cascaded graph algorithms, orchestrated by an agentic AI layer, identify root causes in complex network failures. The approach was demonstrated with NTT DOCOMO at the Mobile World Conference, achieving root cause analysis in minutes on commercial networks. The article traces how graphs evolved from topology models to active reasoning substrates for agents.

AIThe author gave the first half of a children's saying, "if you're serious, it's interesting," to qwen3-32b-instruct. The model's top predictions were "success" and "gain," while "interesting" did not rank in its top twenty, illustrating how LLMs work by predicting the next word.

AIPhysicist Matthew Schwartz describes building BootLoops, an open-source harness for exact quantitative calculations, after choosing problems suited to Claude's strengths. He reports that Claude solved long-standing integrals and found connections across ecology, population genetics, economics, and linguistics, with domain experts steering results toward questions those fields care about. The post states that the approach required constant human oversight, since Claude often overstated results and misjudged time.

Why it matters: The guest post explains why scientists often find current AI tools frustrating and offers a method for finding problems where AI and researchers match, backed by concrete projects.

AIApple researchers introduced RLTL;DR, a reinforcement learning method in which an agent writes its own one-line insight after each failed attempt and learns to map tasks to those insights. On challenging tool-calling and coding datasets filtered to Pass@128 = 0, standard GRPO training of a Qwen 3.5 9B Thinking policy stayed at 0% to 1% Pass@1, while RLTL;DR reached 14–31% with insights in context and 12–13% without them at evaluation. A compact variant, SFTL;DR, trained on just 4k task-insight tuples recovered nearly the full performance of RLTL;DR.

AIAmazon researchers will attend COLM 2026 in San Francisco next week with accepted publications on agentic AI, LLM post-training, multi-agent systems, and automated reasoning. The post links to their research and booth schedule.

AIGoogle AI announced Gemini 4 Argon, a new frontier model built for deep reasoning across long, complex workflows in software engineering, legal and finance knowledge work, and cybersecurity defense. Google says it is expanding the model's output token limit to 1M tokens. Argon is rolling out first to trusted cyber defenders in the Fairwind Program, with broader availability to follow as soon as possible.

Why it matters: The benchmark table compares Gemini 4 Argon against GPT-6 Astra and Claude models across knowledge work, coding, and multimodal tasks, showing where it leads and trails.

AIGoogle DeepMind's Argon supports a 1M token output limit, enabling deeper reasoning on long, multi-step problems in a single pass. Early testers are providing feedback before a broader rollout to developers, enterprises, and consumers.

AIGoogle DeepMind announced Gemini 4 Argon, rolling out first to trusted cyber defenders through its Fairwind Program. Argon will launch at an introductory price of $2 per million input tokens and $10 per million output tokens, with output limits raised to 1M tokens. The post cites a 77.9% score on DeepSWE v1.1 and 91.7% on LVBench, and says broad availability will follow safeguard testing.

Why it matters: The post pairs Argon's benchmark claims with the phased release, pricing, and safeguard details, helping readers weigh its frontier-level capabilities against its access limits.

AIStanford HAI promotes Dan Yamins's October 28 talk on Cognitive NeuroAI, which uses neural networks to explain brain function. The talk covers an effort to build a digital twin of the human brain, part of a large-scale computational models session at the neuroscience-AI intersection.

AIResearchers from Tencent Hy, Fudan University, and Tsinghua University released ExplorationBench, a benchmark that tests whether AI systems can discover hidden rules in executable Alien World sandboxes. Across 10 frontier systems, getting feedback from experiments outperformed thinking alone, with the best run reaching 89.0% after four rounds. The authors note that rankings barely transfer between the two worlds, and the code is listed as coming soon.

AIOpenBMB introduces Diffusion Reward Models (DRM), which learn the full reward distribution of human preferences instead of collapsing them into one scalar score. The approach preserves disagreement and uncertainty, enabling distribution-aware Best-of-N ranking and a new test-time scaling axis by sampling more reward outputs. DRM also improves downstream policy performance over scalar reward baselines when used as the reward in RLHF, according to the post.

AIOllama now supports JEV-style decision models, with documentation for the decision capability and API. The post links to a blog announcement, decision capability docs, and API docs, but gives no further details on the models or their performance.

AIKorean AI lab Upstage has released Solar Mini 4, a proprietary reasoning model that scores 24 on the Artificial Analysis Intelligence Index with 35B total and 3B active parameters. It is priced at $0.10/$0.40 per 1M input/output tokens and has a 1M-token context window, but averages 7.1 minutes per task due to heavy output token use. Its weights are not released, and its size cannot be independently verified.

AIArtificial Analysis reports that Google's Gemini 4 Argon scores 53 on its Intelligence Index with high reasoning, matching GPT-6 Astra (max) and one point ahead of GPT-6.1 Sol (max). At the current 50% launch discount, its cost per task is $1.99, about 60% of GPT-6 Astra's $3.26, but the discount's end date is unconfirmed and standard pricing would raise it to $3.98. The model is being rolled out to selected users and is not publicly available.

Why it matters: The benchmark compares Gemini 4 Argon's cost per task and hallucination rate with GPT-6 Astra, showing where its value depends on a temporary 50% discount.

AINumerical differences between a rollout engine and a trainer can make reinforcement learning collapse even when algorithm and data stay identical. In a GLM 5.2 experiment, reward fell from about 0.9 to under 0.2 around step 20 without alignment, while aligned numerics kept reward stable over 25 steps. A Qwen3.5-MoE investigation traced a significant mismatch to how expert outputs were combined, and router replay alone was judged insufficient.

AIFrançois Chollet suggests a meta-benchmark for human-level generality: an AI system should be able to pass ARC-AGI-(n+1) immediately upon its release. A quoted reply from @dylantmoore adds that such a test would presumably not hold up forever.

AIBAAI released AREX-2, a 27B-parameter long-horizon agent model that improves solutions over multiple test-time rounds by proposing, measuring, reflecting, and revising. It was trained on machine-learning and algorithmic-programming tasks with verifiable feedback, and the source reports that this self-improvement transfers to deep research. The model is Apache License 2.0 licensed and has a 262,144-token context length.

Why it matters: The source compares AREX-2 against closed and open models on coding and deep-research benchmarks, showing how test-time self-improvement is measured across task types.

AIJerry Liu hosted a dinner with Snorkel's Vincent Sun on evals and RL environments, a topic shaped by models rapidly saturating benchmarks. The conversation highlighted that building fair RL environments is hard, since failures are difficult to attribute to input, harness, or reward model, and that long-horizon evals spanning weeks or months remain very difficult. The post also noted that regulated industries still require human-in-the-loop review because 80% accuracy is not sufficient.

AIResearchers from Tsinghua NLP and collaborators show that on-policy distillation with a single training query recovers 87% of full-data gains on math, reaching 68.5 versus 69.8 by step 300. The paper attributes the slow progress to how fast the student absorbs the teacher's signal rather than to dataset size. Code and the paper are publicly available on GitHub and Hugging Face.

Why it matters: The paper isolates training data from the algorithm, showing one query nearly matches full-data on-policy distillation, which reframes where post-training gains come from.

AIModelScope announced IQuest-Q1, a 320B MoE model with 15B active parameters and a 512K context window for agentic coding. The post reports scores of 84.5 on CyberGym, 83.2 on Terminal-Bench 2.1, 64.6 on DeepSWE v1.1, and 63.0 on NL2Repo, and says weights are released under the IQuest-Q1 License.

AIInternLM's AutoVerifier, built on Qwen3_5MoeForConditionalGeneration with about 68 GiB of weights across 40 safetensors shards, evaluates natural-language mathematical proofs, explains errors, and identifies the earliest incorrect step. It serves as the automatic grader for AdvancedMathBench's ProverBench, which accepts a proof only when all eight judgments report -1. The model is a learned grader rather than a formal proof checker and can make errors.