Stanford AI Lab

A Stanford ICML paper finds LLMs predict each other's answers better than the truth, so polling or ensembling multiple models fails to recover factual accuracy the way self-consistency does for math and code.

Stanford's Theory of Space benchmark tests whether foundation models can build, revise and use spatial beliefs through active exploration. Results show a text-versus-vision gap, inefficient exploration, and belief inertia when objects move, across GPT-5.2, Gemini 3 Pro and Claude Sonnet 4.5.

A Stanford/HazyResearch team presents a closed-form method for writing facts directly into a Transformer's MLP layers without gradient descent, a training-free way to inject knowledge into a language model. Accepted at COLM 2026.

TRACE is a self-improvement method where an agent diagnoses the specific capabilities behind its own failures and trains on them; a TRACE-trained Qwen3.6-27B reaches 73.2% on SWE-bench Verified, beating larger models with a quarter of the rollouts GRPO or GEPA need.

A Stanford AI Lab blog post traces the intellectual history of chain-of-thought monitoring, tracking how reading a model's reasoning trace became a central AI safety strategy.

Stanford's R&B-EnCoRe lets a vision-language-action model learn which reasoning steps actually help it act, using self-supervision with no reward model, verifier, or human annotation.

A new technique called Prefix Sliding lets LLMs selectively forget earlier parts of long reasoning traces, avoiding the memory cost of full attention and the accuracy loss of compaction, without any retraining.

Stanford's LLM-as-a-Verifier work shows scaling verification pushes agent performance to state of the art across Terminal-Bench V2, SWE-Bench Verified, RoboRewardBench, and MedAgentBench, and its fine-grained feedback also boosts RL sample efficiency.

New research finds LLMs now outperform dedicated embedding models on retrieval tasks but at substantially higher inference cost, giving a concrete cost/quality tradeoff for choosing between the two approaches.

VAGEN trains vision-language agents with RL to reason explicitly about current state and predicted transitions, using an LLM-judge world-modeling reward and turn-then-token credit assignment. A 3B model reaches 0.82 across five tasks, ahead of GPT-5, Gemini 2.5 Pro and Claude 4.5.

Stanford AI Lab researchers are exploring whether a randomly initialized model can generate all its own training data through self-play alone, with no real-world corpus involved.

TTT-E2E, from Stanford with NVIDIA and Astera, lets models update their weights during deployment by treating in-context data as training data, enabling continual learning from experience rather than fixed weights.

M* is a unified inference runtime for composite multimodal models that matches or beats specialized serving systems, with reported gains up to 2.7x on omni text-to-speech and 12.5x on world-model rollouts.

Auto-psych is a Stanford system where AI agents autonomously generate theories, design experiments, recruit human participants online, and iterate on real results, automating the full research loop.

Stanford's T* treats long-form video understanding as a temporal search problem, letting a model locate the handful of relevant 'needle' frames instead of processing an entire video sequence.

Stanford released Terminal-Bench-Science, a new benchmark for evaluating AI agents on real research workflows across scientific domains, built by the Terminal-Bench team with input from domain experts.

The TTT-Discover paper runs reinforcement learning on one problem at inference using gpt-oss-120b. It reports new state of the art on an Erdős bound, GPUMode TriMul kernels, an AtCoder contest and single-cell denoising, with ablations. It also admits a few misses and caveats.

Stanford SALT's SWE-chat v2 expands to 17,968 real coding agent sessions and about 230K user prompts from public repositories. It includes transcripts, tool calls, subagent tasks, skill invocations and per-commit human versus agent attribution. Gated on Hugging Face.

Distill to Detect surfaces hidden biases in fine-tuned models by distilling the behavioral shift from the base model into a small cartridge, amplifying subliminal preference signals enough for existing auditing tools to catch them.

Stanford's Context-Sharded Block Parallelism (CSBP) is a distributed training strategy for diffusion LLMs delivering up to 7.59x faster speculative-decoding drafter training and 1.61x faster block-diffusion fine-tuning, with gains scaling with context length.

A Stanford ICML oral paper quantifies how residual duplicate pretraining data degrades language models, finding the worst-case repetition pattern is predictable from model size and can waste up to 33% of training compute.

Stanford's Freeform Preference Learning has annotators describe preference axes in natural language instead of picking a single winner, learning axis-conditioned rewards that extract better robot policies.

Stanford researchers flag flawed benchmark questions using response-pattern statistics from measurement theory, then use an LLM judge as a first pass. Across nine benchmarks, up to 84% of flagged questions were confirmed invalid. Code and data are public.

A Stanford-ETH collaboration introduces SDPO, an RL method trained on natural-language feedback like error messages or LLM-judge critiques, reporting large gains over scalar-reward GRPO on very hard tasks.