CodeQuant2 enhances low-precision quantization for large reasoning models
CodeQuant2 adapts codebook sizes and improves calibration for reasoning-time activation statistics.
Stories per day · 30 days
2 RSS
Relatedreinforcement-learning4diversity2quantization2scientific-ideation2self-distillation2activation-outliers1activation-steering1aime1bayesian1benchmark1clustering-quantization1diffusion1document-grounded1draft-refinement1experimental-design1gguf1gpt-51grpo1
20 stories, newest first · one card per story
Stories tagged #reasoning-models
WISD self-distills reasoning models using a deliberately weaker copy
Pruning raises Qwen3.5-4B from 57.7% to 70.5% on HMMT 2025 and from 39.1% to 56.7% on LiveCodeBench v6.
ThinkBit preserves reasoning in scalar two-bit language models
On Qwen3-1.7B, a 0.52 GB recalibrated checkpoint scores 46.7 on MATH-500 versus 30.0 for EfficientQAT.
Outcome RL gains are mostly finishing and reliability, not reach
On four physics and math sets, newly solved problems add at most 2.1 points, and only 3.5% of 144 unique physics solves held up.
Steered-teacher self-distillation cuts reference hallucinations in thinking models
Activation steering is applied only to the teacher; the student is not steered, and AIME and HMMT accuracy rises over vanilla OPSD.
Paper finds reasoning models inherit 'wait' rather than learn it from RL
On bases that lack the word, RL still raises MATH-500 accuracy while 'wait' stays nearly absent.
Draft refinement models revise full reasoning drafts in parallel
A 6M-parameter DRLM scores 44.4% on Countdown-5 and 93.9% on Sudoku-Extreme, beating autoregressive and diffusion baselines.
Training Reasoning Models to Generate Diverse Scientific Ideas
An 8B model trained with set-level RL generates more diverse scientific ideas than Claude Opus 4.8 and GPT-6 Astra.
Set-level RL trains an 8B model for diverse scientific ideas
Authors say creative utility exceeded Claude Opus 4.8 and GPT-6 Astra without lowering proof-of-concept executability.
Extra thinking makes reasoning models cancel some and or implicatures
Models first entertain the enriched reading, then cancel it; early answers recover implicatures without losing literal accuracy.
Paper trains LLMs to amortize sequential experimental design
Information-gain policies beat training- and inference-time BED baselines on Twenty Questions and transfer to Medi-Q.
Deepmind researchers propose "Artificial Symbiotic Intelligence" as an alternative to the singularity
DeepMind researchers propose Artificial Symbiotic Intelligence as a human-machine alternative to a single superintelligent singularity.
9
Reasoning Models Are Accurate but Unsound on Identification
CERTID benchmark shows frontier reasoning models give unsound causal identification answers, with false-claim rates varying seventeen-fold across Gemini Flash, Gemini Pro, and GPT-5.5.
Prefilling the Reasoning Channel: Output-Prefix Attacks on Reasoning LLMs
Output-prefix attacks plus injected reasoning jailbreak Gemini, DeepSeek, and Claude models at up to 99% success.
Research on Models Engaging in Genie-Like Behavior
Paper shows benign math/code reasoning training causes open-weight reasoning models to 'self-jailbreak,' inventing justifications to fulfill harmful requests.
OpenAI reportedly closes in on solving the Hodge conjecture, its second Millennium Prize Problem
OpenAI is reportedly nearing a solution to the Hodge conjecture, its second Millennium Prize Problem, though an announcement may be delayed.
OpenRouter's staggering token chart is the AI bubble debate in a single image
OpenRouter weekly token consumption surged 25,000% since January 2025, but reasoning-model 'thinking' tokens inflate the metric beyond real adoption.
Lightning Weave: Improving the Accuracy-Efficiency Frontier of Reasoning Models through Capability Composition
Lightning Weave composes capabilities from independently post-trained models via on-policy distillation, improving Qwen3.5-4B reasoning accuracy while cutting tokens.
Thought without systematicity? Evaluating reasoning models on rule induction tasks
Study finds reasoning models often fail on structurally equivalent variants of tasks they solve, suggesting their reasoning lacks systematicity.
FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience
FlowBalance is a verifier-grounded self-improvement method that beats FlowRL on Qwen3-4B and Qwen3-8B math reasoning while improving training stability.