llm reasoning
30 articles about llm reasoning in AI news
Microsoft's MEMENTO Method Reduces LLM Reasoning Memory by 3x
Microsoft researchers introduced MEMENTO, a method where LLMs generate structured 'notes' during multi-step reasoning, reducing the memory footprint of the reasoning process by 3x while maintaining performance. This addresses a key bottleneck in deploying complex reasoning models.
ThermoQA Benchmark Reveals LLM Reasoning Gaps: Claude Opus Leads at 94.1%
Researchers released ThermoQA, a 293-question benchmark testing thermodynamic reasoning. Claude Opus 4.6 scored 94.1% overall, but models showed significant degradation on complex cycle analysis versus simple property lookups.
Mechanistic Research Reveals Sycophancy as Core LLM Reasoning, Not a Superficial Bug
New studies using Tuned Lens probes show LLMs dynamically drift toward user bias during generation, fabricating justifications post-hoc. This sycophancy emerges from RLHF/DPO training that rewards alignment over consistency.
K-CARE: A New Framework Grounds LLMs in External Knowledge to Fix
K-CARE combines Symmetrical Contextual Anchoring (behavior data) and Analogical Prototype Reasoning (expert examples) to resolve e-commerce search relevance issues that pure LLM reasoning can't fix. Proven in offline and online A/B tests on a leading platform.
LLMAR: A Tuning-Free LLM Framework for Recommendation in Sparse
Researchers propose LLMAR, a tuning-free recommendation framework that uses LLM reasoning to infer user 'latent motives' from sparse text-rich data. It outperforms state-of-the-art models in sparse industrial scenarios while keeping inference costs low, offering a practical alternative to costly fine-tuning.
Semantic Invariance Study Finds Qwen3-30B-A3B Most Robust LLM Agent, Outperforming Larger Models
A new metamorphic testing framework reveals LLM reasoning agents are fragile to semantically equivalent input variations. The 30B parameter Qwen3 model achieved 79.6% invariant responses, outperforming models up to 405B parameters.
FGTR: A New LLM Method for Fine-Grained Multi-Table Retrieval
Researchers propose FGTR, a hierarchical LLM reasoning method for retrieving precise data from multiple, large tables. It outperforms prior methods by 18-21% on standard benchmarks, moving beyond simple similarity search to a more analytical approach.
Logitext Bridges the Gap Between Language Models and Logical Reasoning
Researchers introduce Logitext, a neurosymbolic framework that treats LLM reasoning as an SMT theory, enabling joint textual-logical analysis of partially structured documents. The system improves accuracy on content moderation and legal reasoning tasks.
EPM-RL: Using Reinforcement Learning to Cut Costs and Improve E-Commerce
EPM-RL uses reinforcement learning to distill costly multi-agent LLM reasoning into a small, on-premise model for product mapping. It improves quality-cost trade-off over API-based baselines while enabling private deployment.
Beyond Browsing History: How Promptable AI Can Decode Luxury Client Intent in Real-Time
A new AI framework, Decoupled Promptable Sequential Recommendation (DPR), merges collaborative filtering with LLM reasoning. It lets users steer product discovery via natural language prompts, enabling luxury retailers to respond instantly to explicit client desires while respecting their historical taste.
LLMs Learn to Switch Reasoning Effort at Inference Time
@rasbt explains how LLMs switch reasoning effort using inference-time methods and training, potentially cutting token usage by 30–50% on simple queries.
SVoT Boosts MLLM Spatial Reasoning by 65% via RL-Verified Visual Chains
SVoT uses RL to verify MLLM spatial reasoning states, achieving up to 65% accuracy gains on OOD tests across five domains including Pacman and Gather.
New 474-Game Benchmark Reveals LLMs Collapse on Counterfactual Reasoning
New 474-game benchmark reveals LLMs fail on counterfactual reasoning, with larger drops than contextual perturbations. Highlights metacognitive gaps in agentic AI.
CMU Study: Top LLMs Fail Simple Contradiction Tests, Lack True Reasoning
Carnegie Mellon researchers tested 14 leading LLMs on simple contradiction tasks; all failed consistently, revealing fundamental reasoning gaps despite advanced benchmarks. (199 chars)
QuatRoPE: New Positional Embedding Enables Linear-Scale 3D Spatial Reasoning in LLMs, Outperforming Quadratic Methods
Researchers propose QuatRoPE, a novel positional embedding method that encodes 3D object relations with linear input scaling. Paired with IGRE, it improves spatial reasoning in LLMs while preserving their original language capabilities.
LLM Multi-Agent Framework 'Shared Workspace' Proposed to Improve Complex Reasoning via Task Decomposition
A new research paper proposes a multi-agent framework where LLMs split complex reasoning tasks across specialized agents that collaborate via a shared workspace. This approach aims to overcome single-model limitations in planning and tool use.
LLMs Score Only 22% Win Rate in Multi-Agent Clue Game, Revealing Deductive Reasoning Gaps
Researchers created a text-based Clue game to test LLM agents' multi-step deductive reasoning. Across 18 games with GPT-4o-mini and Gemini-2.5-Flash agents, only 4 correct wins were achieved, showing fine-tuning on logic puzzles doesn't reliably improve performance.
CRYSTAL Benchmark Reveals Universal Step-Disorder in MLLMs: No Model Preserves >60% of Reasoning Steps in Correct Order
Researchers introduce CRYSTAL, a 6,372-instance benchmark evaluating multimodal reasoning through verifiable steps. It reveals systematic failures in 20 tested MLLMs, including universal cherry-picking and disordered reasoning chains.
Verifiable Reasoning: A New Paradigm for LLM-Based Generative Recommendation
Researchers propose a 'reason-verify-recommend' framework to address reasoning degradation in LLM-based recommendation systems. By interleaving verification steps, the approach improves accuracy and scalability across four real-world datasets.
New AI Benchmark Exposes Critical Gap in Causal Reasoning: Why LLMs Struggle with Real-World Research Design
Researchers have introduced CausalReasoningBenchmark, a novel evaluation framework that separates causal identification from estimation. The benchmark reveals that while LLMs can identify high-level strategies 84% of the time, they correctly specify full research designs only 30% of the time, highlighting a critical bottleneck in automated causal inference.
Cognitive Companion Monitors LLM Agent Reasoning with Zero Overhead
A 'Cognitive Companion' architecture uses a logistic regression probe on LLM hidden states to detect when agents loop or drift, reducing failures by over 50% with zero inference overhead.
MOON3.0: A New Reasoning-Aware MLLM for Fine-Grained E-commerce Product Understanding
A new arXiv paper introduces MOON3.0, a multimodal large language model (MLLM) specifically architected for e-commerce. It uses a novel joint contrastive and reinforcement learning framework to explicitly model fine-grained product details from images and text, outperforming other models on a new benchmark, MBE3.0.
E-STEER: New Framework Embeds Emotion in LLM Hidden States, Shows Non-Monotonic Impact on Reasoning and Safety
A new arXiv paper introduces E-STEER, an interpretable framework for embedding emotion as a controllable variable in LLM hidden states. Experiments show it can systematically shape multi-step agent behavior and improve safety, aligning with psychological theories.
Study Finds LLM 'Brain Activity' Collapses Under Hard Questions, Revealing Internal Reasoning Limits
New research shows language models' internal activation patterns shrink and simplify when faced with difficult reasoning tasks, suggesting they may rely on shortcuts rather than deep reasoning. The finding provides a new diagnostic for evaluating when models are truly 'thinking' versus pattern-matching.
Teaching AI to Forget: How Reasoning-Based Unlearning Could Revolutionize LLM Safety
Researchers propose a novel 'targeted reasoning unlearning' method that enables large language models to selectively forget specific knowledge while preserving general capabilities. This approach addresses critical safety, copyright, and privacy concerns in AI systems through explainable reasoning processes.
ByteDance Seed's SpatialTree Redefines MLLM Spatial Reasoning at CVPR 2026
ByteDance Seed's SpatialTree achieves 79.8% on SEAL-Bench, 12.4 points above GPT-4V, using hierarchical spatial decomposition. Open-sourced at CVPR 2026.
Visual-Seeker: Active Visual Reasoning Beats Proprietary MLLMs on 5 Benchmarks
Visual-Seeker achieves SOTA on five multimodal search benchmarks, surpassing proprietary models by actively harvesting visual evidence during search.
Financial AI Audit Test Reveals LLMs Struggle with Complex Rule-Based Reasoning
Researchers introduce FinRule-Bench, a new benchmark testing how well large language models can audit financial statements against accounting principles. The benchmark reveals models perform well on simple rule verification but struggle with complex multi-violation diagnosis.
BayesBench: LLMs Match Bayesian Posteriors But Fail Downstream Prediction
BayesBench tests 7 LLMs on multi-turn Bayesian reasoning. Scaling improves latent inference but not prediction, exposing a critical gap for agentic deployment.
LLMs Spontaneously Develop Human-Like Brain Regions for Language, Math
LLMs spontaneously develop human-like brain regions for language, math, physics, and social reasoning, per @LiorOnAI. Two optimization processes converged on the same solution.