A new arXiv paper (2610.09186v1) argues that step-by-step reasoning and pattern recognition in large language models lie on a spectrum defined by how much preceding context each next token depends on,…
Training Qwen3-4B with counterexample-only supervised fine-tuning collapsed true-theorem recognition from 0.27 to 0.00, while reinforcement learning with a sparse outcome-only reward repaired the coll…
Researchers introduced Sigma, described as the first large-scale continuous diffusion language model at 3B and 8B parameters, built on steerable low-dimensional ODE/SDE latent trajectories and trained…
A Mixture-of-Agents (MoA) study measuring per-token sufficient compute found that a 0.5B agent reproduces 92–95% of reference tokens on three core benchmarks, while the most expensive 10% of tokens ac…
GoldiMask, a fine-tuning method for discrete diffusion language models introduced in arXiv paper 2609.38385v1, selects which response tokens to reveal as context by approximately maximizing a submodul…
A controlled evaluation of eight generated LLM harnesses plus a baseline and nine byte-identical baseline copies on 386 MATH-500 tasks found that identical programs alone yield 2.16 percentage points …
ELF-REG, a method that scales Embedded Language Flows (ELF) continuous diffusion language models to reasoning tasks, reached 55.96% pass@1 on GSM8K at 64 network function evaluations and 13.39% on MAT…
A reasoning-aware compression framework that selectively restores the most quantization-sensitive circuits to FP16 achieves Pareto-optimal energy-accuracy points unreachable by uniform quantization, a…
Researchers introduced AdaptiveSpec, a training-free per-step speculative decoding method that adapts both draft verification and tree shape from internal decoding signals, improving throughput over E…
A new arXiv study (2608.20362v1) finds that exact-match verifiers in reinforcement learning with verifiable rewards (RLVR) introduce language-dependent false-negative reward noise in multilingual sett…
Researchers propose DUET, a token-selective on-policy distillation method that uses two identical-weight teachers differing only in prohibition visibility to improve LLM compliance with runtime-inject…
Researchers introduced Attention-Aware Transform Coding (AATC), a KV cache compression method that reduces memory use by approximately 5.8x while maintaining near-lossless accuracy on models like Llam…
Researchers propose Thought-Aware Attention Matching (TAM), a KV cache compaction method that segments chain-of-thought reasoning into blocks, allocates compression budgets adaptively, and protects pi…
At YCML, YC's first machine learning research showcase at Startup School 2026, eight researchers presented papers on topics including model reasoning, formal mathematics, video agents, and robotics. N…
LOCKS, a new method from arXiv, enables efficient long-context decoding by giving each page of the key-value cache its own compact spectral summary, reconstructing within-page logits, and attending on…
Kimi K3 is 30x cheaper than GPT-5 for output tokens while leading the LMArena leaderboard with 1,289 ELO, according to a 2026 comparison. For a customer support bot handling 10M input and 5M output to…
Researchers propose LISA (Linear-Indexed Sparse Attention), a plug-and-play attention module that reduces inference complexity from O(n^2) to O(nM) for long-context reasoning models. Experiments on De…
Researchers propose BIRD (Bootstrapped Iterative Self-Reasoning Distillation), a two-stage method that improves compressed reasoning by first generating concise, correct traces from a base model under…
SPARK, a new approach to diagnosing hidden-state failures in AI reasoning, boosted accuracy on the MATH-500 benchmark for Qwen3-4B from 82.0% to 84.6% and for Qwen3-8B from 82.4% to 85.6%, according t…
A new study evaluating 67 frontier AI models from 21 providers reveals that enterprises underestimate AI model failure rates by 2.25 to 2.5 times due to a mathematical limit called the "co-failure cei…