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 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 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,…
The Role-Context-Task-Constraint (RCTC) model is the most effective framework for structured prompting, eliminating ambiguity by assigning a professional persona, providing context, specifying an impe…
Researchers propose Moir, a method that estimates the preservation covariance directly from a language model's own decoding distribution to prevent capability erosion during knowledge editing. In test…
In a head-to-head test of eight GSM8K math word problems, gpt-oss-120b and Kimi K3 tied on seven tasks, both returning the correct numeric answer every time. The only separation came on the eighth tas…
Researchers propose PPL-Factory, a task-aware and budget-aware data selection framework for large language model fine-tuning that uses perplexity-based scores to select informative training samples. E…
Researchers introduced PPO-HSC (Proximal Policy Optimization with High-order Sampling Coverage), an exploratory reinforcement learning framework that addresses mode collapse in Large Language Model fi…
A new paper from researchers systematically evaluating widely used AI safety datasets AdvBench and HarmBench finds that these benchmarks contain contrived data points with excessive triggering cues an…
Researchers from University of California, Santa Barbara, Carnegie Mellon University, LMSYS Org, and University of Wisconsin–Madison introduced the Length Value Model (LenVM), a token-level framework …
A developer known as flirp spent 21 GPU-hours training Qwen2.5-1.5B-Instruct to reason in latent space without decoding tokens, but the model's hidden-state reasoning collapsed into content-free attra…
A new paper on arXiv (2607.11937v1) introduces viable path entropy (VPE), a finite-budget measure of verified continuation capacity for intelligent systems under repeated reflection. In language-model…
A new study from arXiv reveals that post-training quantization can silently alter how large language models reason even when task accuracy is preserved. Analyzing 30,000 chain-of-thought outputs from …
Researchers introduce KV-PRM, a process reward model that reduces scoring cost from O(L²) to O(L) by reading the KV cache from LLM generation instead of re-encoding trajectory text, achieving up to 5,…