AI Reasoning News and AI Updates

Latest AI Reasoning news, research, benchmarks, product releases, and industry adoption updates.

Latest Articles

• LFM2.5 Q4\_0 Checkpoints from Quantization-Aware Distillation — Liquid AI releases QAD Q4_0 GGUF checkpoints for four LFM2.5 models (230M, 350M, 1.2B, 2.6B), recovering ~97% of BF16 accuracy at native 4-bit speed and memory — making high-quality edge inference practical without larger quantization formats.
• LLM Enhancement with Domain Expert Mental Model to Reduce LLM Hallucination with Causal Prompt Engineering — A causal prompt engineering framework addresses LLM hallucination from tacit expert knowledge — knowledge never written down and unavailable to RAG or knowledge graphs. By encoding domain expert decision logic as Expert Mental Models using monotone Boolean function theory, it supplies structured reasoning that retrieval methods structurally cannot.
• MCTS-KBQA: Monte Carlo Tree Search with Information Gain Rewards for Knowledge Base Question Answering — Fast MCTS replaces expensive rollout-based search in knowledge base question answering with a training-free information gain reward computed by a single LLM forward pass — consistently improving the accuracy-cost tradeoff on four KBQA benchmarks. Accepted at CIKM 2026.
• SGHA: Evidence-Grounded Research Problem Discovery with Local Language Models — SGHA discovers scientific research gaps using only a local 9B open-weight model and a structured evidence graph — no proprietary API calls — producing traceable research problems with assumptions, objectives, and success criteria, compared favorably against AI Scientist-v2 across 5 machine-learning domains.
• Neuro-symbolic learning over OWL 2 DL via consequence-based compilation to differentiable circuits — Baobab compiles full OWL 2 DL (SROIQ) ontologies into differentiable Sentential Decision Diagrams for neuro-symbolic learning — the first NeSy system to handle the complete description logic without restricting to the Horn fragment, with soundness machine-checked in Lean 4, accepted at NeSy 2026.
• The Price of Thinking: Reasoning Effort as a Model-Specific API Contract — A preregistered study finds that explicitly requesting "high effort" from Sonnet 5 costs $0.01 more per AIME 2026 call but produces no statistically detectable accuracy improvement — the confidence interval allows up to a 4.67 pp gain but cannot confirm one.
• The Problem Is the Problem: Towards Scalable Mathematical Discovery — The FAR pipeline scanned 5,245 combinatorics papers, filtered them to 4,717 open conjectures, and surfaced 77 items for human review — among which the team confirmed real mathematical discoveries on longstanding open problems from Erdős-Straus and others.
• Explicit State Elicitation Is Not Enough: A Controlled Audit of Memory-Policy Classification — A controlled audit finds that prompting LLMs to output explicit memory-policy state classifications does not significantly improve routing accuracy for Llama-3.3-70B or GPT-OSS-120B — and that standard example-level accuracy metrics overstate counterfactual consistency.
• Beyond the Trace: Coupling an Interpretable Reasoning-State Readout to Native MoE Routing — J64 and R64 metrics expose MoE routing behavior as interpretable readouts of chain-of-thought reasoning depth, with routing-only proxies that preserve 0.9-3.2 accuracy points versus full model runs on math, coding, and science benchmarks.
• Graph Surgery and the Do-Operator: A Precise Correspondence for Acyclic Structural Causal Models — A formal proof establishes that the do-operator in Pearl's causal calculus is exactly equivalent to graph surgery on acyclic structural causal models, providing a clean mathematical foundation linking the two most-used formalisms in causal AI research.