# large language model

3 related articles

Subscribe to updates

MIT Technology Review Oct 2, 2026
AlphaGo team member Thore Graepel leaves DeepMind, arguing AlphaGo-style search—not LLM chain of thought—is the path to genuine machine reasoning
Using move 37 from the 2016 Seoul match, AlphaGo core member Thore Graepel argues the move came from AlphaGo's search machinery rather than intuition, contends that today's large language models—relying on next-token prediction and chain of thought—do not reason in a scientist's sense, and proposes borrowing AlphaGo's game-tree structure so that general reasoning systems maintain an explicit, inspectable epistemic state updated by evidence, with an independent component judging each step by how much it actually resolves uncertainty.
Using move 37 from the 2016 Seoul match, AlphaGo core member Thore Graepel argues the move came from AlphaGo's search machinery rather than intuition, contends that today's large language models—relying on next-token prediction and chain of thought—do not reason in a scientist's sense, and proposes borrowing AlphaGo's game-tree structure so that general reasoning systems maintain an explicit, inspectable epistemic state updated by evidence, with an independent component judging each step by how much it actually resolves uncertainty.
Machine Reasoning AlphaGo large language model Chain-of-Thought
arXiv Sep 26, 2026
Adjusting a large model's logits at inference with a pair of smaller specialized models removes look-ahead bias in financial prediction without retraining the frontier model
Addressing the look-ahead bias that arises when LLMs are applied to financial predictive tasks because they were trained on long time-series data, and the prohibitive cost of retraining frontier models from scratch with a specific knowledge cutoff, the work introduces a fast, effective, and low-cost alternative that guides generation at inference time by adjusting the logits of a large base model using a pair of smaller specialized models -- one fine-tuned on information to be forgotten and one on information to be retained -- and reports that the method effectively removes both verbatim and semantic knowledge, corrects biases, and outperforms prior methods.
Addressing the look-ahead bias that arises when LLMs are applied to financial predictive tasks because they were trained on long time-series data, and the prohibitive cost of retraining frontier models from scratch with a specific knowledge cutoff, the work introduces a fast, effective, and low-cost alternative that guides generation at inference time by adjusting the logits of a large base model using a pair of smaller specialized models -- one fine-tuned on information to be forgotten and one on information to be retained -- and reports that the method effectively removes both verbatim and semantic knowledge, corrects biases, and outperforms prior methods.
University of Chicago
inference-time logit adjustment look-ahead bias mitigation knowledge unlearning large language model
arXiv Sep 27, 2026
SkillGym turns human skills into verifiable training environments, letting a 35B model reach 51.47% on SkillsBench and exceed reported scores of Claude Sonnet 4.6 and others
SkillGym introduces a framework that transforms human-written agent skills into executable, verifiable training environments: its skill-to-task pipeline instantiates concrete tasks, verifies outcomes with code-based checkers, and assesses empirical skill dependence through contrastive executions; it constructs and releases 2,756 environments across 12 categories and collects 8,364 successful trajectories (averaging 49 tool calls and over 60k logged text tokens), supporting supervised fine-tuning on verified workflows and reinforcement learning with outcome-based rewards; under Claude Code, supervised fine-tuning improves Qwen3.5-35B-A3B by 199 Elo on GDPval-AA v2, 19.10 percentage points on Terminal-Bench 2.1, and 28.13 and 12.38 points on SkillsBench v1.
SkillGym introduces a framework that transforms human-written agent skills into executable, verifiable training environments: its skill-to-task pipeline instantiates concrete tasks, verifies outcomes with code-based checkers, and assesses empirical skill dependence through contrastive executions; it constructs and releases 2,756 environments across 12 categories and collects 8,364 successful trajectories (averaging 49 tool calls and over 60k logged text tokens), supporting supervised fine-tuning on verified workflows and reinforcement learning with outcome-based rewards; under Claude Code, supervised fine-tuning improves Qwen3.5-35B-A3B by 199 Elo on GDPval-AA v2, 19.10 percentage points on Terminal-Bench 2.1, and 28.13 and 12.38 points on SkillsBench v1.
East China Normal University
Zhilong Ge（First author）
SkillGym Agent Skills Supervised Fine-tuning Reinforcement Learning large language model

You’re all caught up

AlphaGo team member Thore Graepel leaves DeepMind, arguing AlphaGo-style search—not LLM chain of thought—is the path to genuine machine reasoning

Adjusting a large model's logits at inference with a pair of smaller specialized models removes look-ahead bias in financial prediction without retraining the frontier model

SkillGym turns human skills into verifiable training environments, letting a 35B model reach 51.47% on SkillsBench and exceed reported scores of Claude Sonnet 4.6 and others