llm-anthropic 0.31 adds Anthropic Fable 5.2 and Mythos 5.2
Willison says max reasoning on Fable 5.2 spent 128,000 tokens thinking and returned no image, a bug he also saw on Opus 5.5.
Stories per day · 30 days
10 RSS
Relatedqwen34reinforcement-learning4inference3language-models3self-distillation3benchmark2chain-of-thought2distillation2mixture-of-experts2ppo2speculative-decoding2vaspo2adversarial1agents1alphago1anthropic1attck1causalcot1
114 stories, newest first · one card per story
Stories tagged #reasoning
StepBack resamples errors for efficient test-time reasoning
StepBack matches self-consistency accuracy while reducing token usage by up to 41.26%.
TOPSD limits teacher supervision to verified prefixes in distillation
It beats conventional on-policy self-distillation on math reasoning in ten languages, especially lower-resource ones.
MEvince Knowledge-Constrained Framework Identifies Malware ATT&CK Techniques
MEvince uses shared behavioral representations and evidence-aligned retrieval to improve exact-ID F1 by 17.4% on 610 malware reports.
Visual perturbations alter VLA actions without changing text rationale
TNAS-guided attacks induce collisions in 52.6% of driving scenes while keeping rationales word-for-word identical.
Neuro-symbolic framework solves Zebra puzzles from natural language
Combines LLM semantic parsing with symbolic exact cover solver for reliable logic puzzle solving.
HydroReasoner advances floodwater detection with knowledge-guided reasoning
HydroReasoner uses Qwen2.5-VL and a floodwater knowledge graph to improve reasoning and generalize to varied scenes while running efficiently on edge devices.
PIVOT mixes teacher corrections into RLVR student rollouts
Mean avg@8 on hard logic games rose from 19.5% for the best baseline to 35.6%.
Calibrated distillation consolidates specialists in dynamic MoE models
On Turing-20B-A2B the method scores 0.9988 across four domains and beats Qwen3-8B on math and code.
Ex-DeepMind researcher proposes AlphaGo-style ledger against hallucinations
Metis Reasoning’s white paper is only a research program, with no working system, benchmarks, or model comparisons.
1
PHYRD benchmark tests LLMs on reproducing and extending physics
Across five models, Reproduce accuracy was 49.21%–68.25% and closed-book Discover was 18.60%–74.42%.
New pruning method balances quality and diversity for reasoning models
Direct pass@k optimization selects partial solutions to maximize correct final answers while penalizing redundancy in test-time scaling.
CausalCoT isolates sub-questions to limit chain-of-thought attention leakage
On reasoning benchmarks, causal sensitivity of intermediate steps rose 12 points on average and up to 21.
Minotaur claims LLM-style reasoning without a language model
The author says one network operation on a single CPU core answers in about a quarter of a second.
Frontier models scored on LIBERO-10 robot tasks without fine-tuning
Across 1,800 episodes, GPT-6 Astra led at 44% success and Gemini-4 Argon followed at 43%.
Rationale contrastive decoding reduces CoT query shortcuts
The training-free method improved answer accuracy on five reasoning benchmarks and made answers depend more on the rationale.
ATEST adapts LLM test-time sampling to live GPU energy
On four A100 datasets, ATEST-Online stayed inside the energy budget and tied or led accuracy as load rose.
ECHO folds multi-token prediction into the backbone itself
Dynamic Depth Sampling draws training depth with decaying probabilities, at half the overhead of training every depth.
Lossy speculative decoding inflates completion length
At the most relaxed settings tested, models accepted 1.34× tokens per round but wrote 1.54× longer completions.
-oOdUn61bDRSykk4H951SoDOHzqoOe1.png&w=828&q=75)
LFM2.5-24B-A2B: Built for fast on-device agents
Liquid AI released open-weight LFM2.5-24B-A2B, a 24B MoE with 2B active parameters for on-device agents.
Waypoint-OPD: Transferring RL-Induced Policy Shifts via Gated On-Policy Distillation
Waypoint-OPD gates RL policy-shift distillation and leads math and code scores for Qwen3 students.
HeirCommit and EvoHeir filter updates in self-evolving language models
HeirCommit cut harmful accepted updates 68-83% versus random inheritance; EvoHeir raised the math average 5.3% over Agent0.
VASPO: Self-Distillation with Privileged Value Estimation
VASPO combines privileged self-distillation with PPO, gaining up to 2.59 points over PPO on math benchmarks.
VASPO: Self-Distillation with Privileged Value Estimation
VASPO blends a privileged self-teacher with PPO, improving math accuracy on Qwen3-4B and OLMo3-7B.