Estimates a model's probability of correctness by leveraging its past graded episodes: XConf recalls similar past cases and prompts the model to reflect and restate confidence, improving calibration and discrimination across reasoning, coding, multimodal QA, and agent tasks at much lower generation cost.
Describes DeepSeek-V4.1-Flash, a 552B-parameter multimodal Mixture-of-Experts LLM that supports up to 1M-token contexts and reduces KV cache footprint using Compressed Sparse Attention 2, FP4 KV caching, and SWA Bounded Replay to lower HBM/SSD and prefill costs.
Analyzes why on-policy distillation can produce excessively long student rollouts by identifying termination-token mismatch between students and teachers, and proposes aggregating functionally equivalent EOS tokens as a shared semantic stopping action to mitigate length inflation.
Provides an open family of K–12 foundation models and a capability-oriented instruction-tuning corpus that teaches models to solve curriculum problems, map items to curriculum concepts, diagnose learner errors, and produce scaffolded pedagogical responses.
Adds token-conditioned quantum residual branches to a frozen masked-diffusion language model: a lightweight hypernetwork emits continuous quantum-circuit coordinates per token, executes a shared sparse IQP-style circuit, and injects classically-expressible expectation readouts back into transformer blocks. Trains only the added branches, scaling to 16–64 qubits with analytic, linear-cost readouts.
Edits an LLM agent's internal task state and action continuations to remove unsupported assumptions and noisy plans, improving downstream decisions instead of reconstructing high-entropy tool outputs. Combines Action Judge, State Revision, and EditAct and shows consistent gains across benchmarks and agent backbones.
Shows that Transformer LLMs output a superposition of next-token distributions when token embeddings from distinct text streams are linearly combined, finds this property is architectural but weakened by pretraining, and demonstrates it can be restored by lightweight fine-tuning and disentangled at decode time.
Shows that private post-training changes leave measurable "behavioral shadows" in task-unrelated single-word outputs and proposes Active Taskless Distillation (ATD) to transfer capabilities to a student using only those single-word teacher responses, without teacher logits or parameters.
Reduces self-attention complexity to O(N log N) by using a coarse-to-fine (pyramid) Top-K block selection with LogSumExp scoring, implemented with hardware-aware Triton kernels for fused routing and scoring—aimed at long-context LMs and retrieval tasks.
Evaluates multimodal memory in large audio language models by testing four acoustic evidence types and four memory operations across multi-session spoken histories, stratified across 8K–64K token context budgets.
Investigates how rollout policy, token-level KL direction, and learning rate each affect LLM distillation across Llama3 and Qwen2.5 on reasoning tasks; finds KL direction and learning rate dominate outcomes while rollout policy has a modest effect.
Teaches small reasoning models to diagnose when additional internal thinking is insufficient and to selectively query stronger models; introduces FlyBy, a supervised + cost-aware RL framework that learns when/what to ask. Improves pass@ metrics on hard benchmarks while reducing serving cost.