Inference efficiency
Generate text by drafting tokens from early layers of the same LLM (via an early-exit point), then verify those drafted tokens using deeper layers. Intermediate-layer logits are made usable by unembedding through the LM head, enabled through a training recipe that teaches earlier layers to predict next tokens.
AI summary
Retrieval & memory
Train and deploy a single-vector embedding model that maps visual page screenshots into dense vectors, using multilingual image-query pairs so that users can retrieve visual documents across languages without OCR or document chunking.
AI summary
Inference efficiency
Evaluate an LLM/agent for multi-step data analysis by requiring it to use an execution environment and answer factoid-style outputs after iteratively reasoning over both structured datasets (e.g., CSV/JSON) and unstructured reference materials (e.g., manuals). Score outputs with objective, format-flexible comparisons (e.g., numeric tolerance, normalized/fuzzy string matching, element-wise list matching) to reduce reliance on subjective grading.
AI summary
Training & fine-tuning
FAST (Frequency-space Action Sequence Tokenization) tokenizes continuous robot action sequences by converting each action dimension from time to frequency using the Discrete Cosine Transform (DCT), sparsifying/rounding insignificant coefficients, then flattening them into a discrete token sequence. Byte Pair Encoding (BPE) is applied to reduce redundancy further, and the process is described as invertible so actions can be reconstructed from tokens.
AI summary
Reasoning & planning
Instrument an agent so each step (inputs, tool calls, and outputs) is traced to a centralized UI, then run targeted evaluations over the captured tool executions using an LLM-as-a-judge to score properties like retrieval relevance.
AI summary
Agents & tool use
Connect to one or more MCP servers, discover their exposed tools, pass those tools to an LLM via the tools field, then run a control loop that alternates between LLM tool calls and executing those tool calls through the MCP client until the LLM stops requesting tools.
AI summary
Agents & tool use
Build an agent as a conversation loop that uses an MCP client to (1) connect to one or more MCP servers, (2) discover and register tools in an LLM-compatible tool schema, (3) stream a single LLM turn with tool definitions, and (4) execute any tool calls requested by the model by routing them to the corresponding MCP server, then feed tool results back into the conversation until an explicit exit condition is reached.
AI summary
Robotics & embodied AI
Run a visuomotor policy in an asynchronous execution loop by decoupling robot action execution from the computation of the next predicted action chunk. While the robot executes the current chunk, the system sends the latest observation to a (possibly remote) policy server to request the next chunk, then fuses overlapping actions from successive chunks using a simple merge rule to reduce jitter.
AI summary
Agents & tool use
A web-native GUI automation approach that decomposes an agent into three independent components: a Policy model that plans and selects actions, a Localizer model that interprets visual UI state to identify precise interaction targets, and a Validator model that checks whether the intended task has been completed successfully.
AI summary
Training & fine-tuning
Train a diffusion transformer using Low-Rank Adaptation (LoRA) while loading the base model in a reduced-bit quantized form (described as typically 4-bit). Only the LoRA adapter weights are updated; the text encoders and VAE are kept frozen, and several memory-reduction steps are used to fit the training within limited GPU memory.
AI summary
Agents & tool use
Run a full desktop (GUI) session inside an isolated Docker container and provide programmatic and/or protocol-based interfaces so an AI agent can observe the screen, launch applications, interact with the UI, and manage files within that sandbox.
AI summary
Reasoning & planning
A trainable, agentic search procedure for formal theorem proving that uses reinforcement-learning to organize, score, filter, and recursively compose intermediate lemmas at test time. It builds proof attempts by sampling lemma-augmented inputs, prunes lemmas that repeatedly fail, recursively generates and integrates sublemmas when proofs fail, and enforces logical soundness by discarding newly generated lemmas whose negations are provable.
AI summary