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…
A new arXiv paper introduces InflationAgent, a four-stage router that accounts for token inflation—the ratio of true workflow cost to single-call cost—which can exceed 2x on difficult tasks. The syste…
A new preprint by N. Trillard, 'Steer on a Sphere: Geometric Control of Transformer Outputs,' shows that a single tangent step on the sphere of hidden states reaches 91–98% of tokens at rank 1 across …
A developer's guide outlines the three pillars for building production-ready AI agents: benchmarking, cost optimization, and tooling. It emphasizes task-level evaluation over static benchmarks, using …
A new study introducing BenchDrift shows that rephrasing benchmark problems while keeping meaning and answer fixed flips LLM correctness in both directions across eight models and three benchmarks (GS…
Researchers propose CORA-Diff, a training-free method that uses native confidence and persistence signals to skip unnecessary denoising steps in diffusion language models, achieving up to 3.32x speedu…
Researchers propose SAPO, a segment-level automatic prompt optimization method that decomposes prompts into role, context, tasks, and output format, achieving the best average score across SQuADv2, Tw…
Trunchbull, a browser-based tool for running LLM benchmarks, lets users author tests and evaluate models in real time, with native support for the harbor authoring system and the Vercel AI SDK. It inc…
AllenAI's Open Instruct framework was used to build an end-to-end post-training pipeline for a compact instruction-tuned language model, integrating Supervised Fine-Tuning, Direct Preference Optimizat…
A new arXiv study (2608.09942v1) finds that chain-of-thought (CoT) prompting is not a universal enhancer of large language model reasoning, but rather a bandwidth bypass that helps only when serial co…
Researchers introduce CODS, a critic-guided data selection method for reusable offline reinforcement learning, which at a 10% budget retains 96.6% of eligible-pool performance across 20 D4RL task-algo…
Researchers introduced Prompt Embedding Probes (PEP), a white-box method for answer-level hallucination detection in large language models that augments standard linear probes with learnable prompt em…
Researchers propose a new contextual bandit algorithm that explicitly models reward decay to improve iterative refinement in Large Language Models (LLMs), addressing over-exploitation from static prom…
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…
A new analysis argues that increasing model size alone cannot solve reasoning limitations in large language models, as adding parameters primarily improves pattern memorization rather than logical und…
A new arXiv preprint (2608.01014v1) introduces Cloud-ScPO, a topology-guided preference-mining framework that derives preference supervision from the hidden-state geometry of large language models (LL…
Researchers introduced Task-Aware Prompt Rewriter (TAPR), a model that reformulates user prompts into task-optimized prompts to improve downstream LLM performance, trained via reinforcement learning w…
A study at 120B scale finds that inserting constitutional content during midtraining produces durable alignment gains, with constitutionally midtrained models outperforming a control on alignment gene…
A developer advising on training an LLM for adaptive questioning recommends keeping difficulty adaptation in application code rather than fine-tuning, using explicit state rendering in system prompts,…