Generates outcome-specific, dialectical rationales with an LLM and derives continuous, calibrated risk scores for irregularly sampled medical time series—mitigating risk polarization. Reports +3.3% average AUPRC and 81% reduction in calibration error across three benchmarks; code released.
AI Agent Papers2026
Guides LLM-based agents to decompose long-horizon research problems and delegate subtasks to constrained subagents, then fine-tunes models on harness-generated trajectories so delegation decisions become internalized. Reports SearchSwarm-30B-A3B achieving top BrowseComp scores for its scale.
AI Agent Papers2026
Lets a single LLM simultaneously act as agent and environment to bootstrap co-evolutional training — using state-prediction process rewards (World-In-Agent) and failure-mode retrieval (Agent-In-World) to reshape training data; reports ~4% average benchmark gain.
Shifts branching and credit assignment in agentic RL from coarse units to fine-grained decision points in generated sequences. Uses a Branching Score combining token uncertainty and policy-induced likelihood gains plus procedure-level advantage scaling; improves performance across 13 benchmarks while keeping efficient tool calls.
AI Agent Papers2026
Survey of methods for engineering interactive environments for LLM-based agents, covering environment modeling, symbolic and neural synthesis, evaluation, and agent–environment co-evolution. Identifies evolution paradigms and future directions like Environment-as-a-Service and multi-agent systems.
Proposes a router redesign for Mixture-of-Experts (MoE) that aligns each router row with its expert's principal singular direction using Manifold Power Iteration (MPI), improving token–expert affinity. MPI applies a 'power‑then‑retract' step to push router rows toward principal singular vectors while enforcing norm constraints; the paper gives convergence theory and pretraining results on 1B–11B MoE models.
AI Agent Papers2026
Benchmarks evolving environments as sequences of progressive updates and introduces EvoMem, a patch-based memory that records structured update histories so LLM agents can reason about environment evolution. Demonstrates measurable gains on EvoArena and other benchmarks.
Implements a blockwise sparse attention (MiniMax Sparse Attention) that scores and Top-k selects key-value blocks per Grouped Query Attention group to enable attention over million-token contexts. Paired with an exp-free Top-k GPU kernel and KV-outer sparse execution, it reduces per-token attention compute and yields large prefill/decoding speedups.
Applies a population-level test-time scaling strategy that uses one model as generator, verifier, refiner, and ranker to search over candidate proofs. Combines generative-verifier RL and a low false-positive verifier with tournament selection to reach competition-level performance on IMO and USAMO.
Uses Parallel Looped Transformers (PLT) to make loop count a practical knob for code models, finding two loops give the best test-time gains. Trains 7B models on 18T tokens and attributes saturation beyond two loops to a gain–cost tradeoff from positional mismatch.
Proposes ZPPO, a distillation method that keeps the teacher inside prompts rather than injecting teacher gradients, using binary- and negative-candidate prompts plus a prompt replay buffer to recover learning signal on hard examples; shows gains for small Qwen3.5 students across 31 multimodal benchmarks.
Proposes Monotonic Inference Policy Improvement (MIPI) and a two-step Monotonic Inference Policy Update (MIPU) to address training–inference probability mismatch in LLM reinforcement learning by constructing sampler-referenced candidate updates and accepting synchronized updates using an inference-gap proxy; shows improved reasoning accuracy and stability under FP8-quantized rollouts.