Researchers propose Gradient-Aligned Reward (GAR), a dense reward method for reinforcement learning from verifiable rewards (RLVR) that uses cosine similarity between rollout and expert-anchor gradien…
A developer testing four language models to improve a self-rewriting AI agent found that none produced promotable edits, despite resolving dozens of bugs. The models, ranging from Qwen3-4B to Mistral …
Palo Alto Networks' Unit 42 research introduces perturbation probing, a diagnostic that identifies the minimal set of neurons responsible for LLM safety refusal behavior, finding that just 50 out of 3…
Researchers introduced Influence-Aware Policy Optimization (IAPO), a method that uses typed influence-dependency graphs to redistribute trajectory-level advantages for credit assignment in multi-turn …
Researchers introduced computation-conditioned credit transport (CCT) and its concrete algorithm CompPO, which uses the Transformer policy's internal attention to parameterize credit assignment in lar…
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…
Xyntetik Runner now trains LoRA adapters directly through quantized GGUF serving weights, eliminating the separate FP16 training copy. In a reproducible experiment, a tool-use LoRA trained on a frozen…
A new study from arXiv (2608.18106v1) finds that Qwen3-4B, a large language model, exhibits overconfidence, especially when prompted to output numeric confidence scores. The researchers identified tra…
A new arXiv preprint (2608.18379v1) introduces a candidate-free control for single-pass test-time aggregation, finding that on AIME-2025 and HMMT-2025 with Qwen3-4B, conditioning on a candidate pool i…
Inco AI released DFlash 2, an updated version of its parallel speculative-decoding method, reporting a 21% increase in average accepted token sequence length over the original DFlash, with only 1.3% a…
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…
Researchers introduced PAIR (Pairwise-Aware Inclusion Reweighting), a method that treats rollout prefixes as vertices and pair-gradient terms as edges of a contrast graph to reduce compute in reinforc…
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…
Researchers propose SAF, a Stable Advantage Fusion framework that combines reinforcement learning with verifiable rewards (RLVR) and on-policy distillation (OPD) for training language models, addressi…
A ServiceNow research project called MosaicLeaks found that AI research agents can leak private documents through their web search queries, even without uploading the files. In tests with the Qwen3-4B…
Anthropic's Jacobian lens, introduced in a July 6 paper, reveals a low-dimensional 'J-space' inside language models that encodes tokens the model can verbalize. Developer Anyesh repurposed this signal…
OpenClaw, the most widely used agent harness, can now be paired with QVAC, an open-source SDK from Tether, to run a fully local, private AI assistant. The setup uses QVAC's OpenAI-compatible server to…
Researchers propose Multi-Agent Protocol Distillation (MAPD), a framework that distills reasoning from proprietary large language models into open-source search agents using a structured JSON protocol…
A new study from a BlueDot Project cohort introduces value dynamics, a framework for measuring, forecasting, and steering how AI values change in self-training loops. Using tools from population genet…
Senbonzakura, a new open-source tool, removes refusal behavior from open-weight language models by cutting multiple refusal directions in activation space at once, achieving zero hard refusals on Qwen…