论文库

When an LLM agent fails, repair methods choose which action to replace and what to put in its place. They often score candidate repairs with greedy replay, a stronger model, an environment oracle, or a judge that reads the failed log. We ask whether these evaluators pick repairs that help the agent that will actually execute them. Holding failed trajectories and candidate single-step repairs fixed, we label every candidate with independent rollouts of the deployed policy and measure how much deployment value each evaluator's choice loses. On failures not screened for repair value, pre-specified tests show that greedy replay loses 0.08-0.13 in expected net success gain in ALFWorld and 0.06-0.13 in net reward in WebShop, relative to selection with the deployed agent's own rollouts. At an equal rollout budget, selecting repairs for Qwen2.5-7B with Qwen2.5-14B rollouts loses 0.25 in ALFWorld and 0.07 in WebShop, twenty 14B rollouts per candidate are worth no more than two of the 7B agent's own, and fresh rollouts confirm that the two models prefer opposite repairs on selected disagreement cases. The agent's own fine-tuned successor is a worse evaluator still: after one round of training on verified repairs, its rollouts lose 0.55-0.60 when they select repairs for the agent it was trained from, as a pre-specified test confirms. Relative to what selection is worth, substitute models forgo 11-53%; on $τ^2$-bench, where selecting is worth only 0.10, they lose 0.03. Repairs should be verified with the executor that will run them, and repair labels should name their executor, including its training version.

Most AI assistants respond only when prompted, remaining silent even when events in the video stream offer meaningful opportunities for conversation. To move beyond purely user-driven interaction, we present AURA 2.0, a streaming video large language model that introduces three core upgrades to AURA. The key advance is Model-Initiated Conversation (MIC), which enables the model to identify topics worth discussing in the evolving visual stream and initiate conversations at appropriate moments without a user query. Stream-Aware Tool Calling (SATC) enables the model to invoke external tools using the current visual and conversational context and incorporate returned information into the ongoing interaction. Interactive Visual Grounding (IVG) supports visual grounding in both user inputs and model outputs. The model can identify objects marked by users on video frames and produce frame-aligned bounding boxes to link its responses to specific objects and regions. To develop these capabilities, we construct a dedicated streaming interaction dataset and adopt a three-stage training scheme to integrate the new behaviors with AURA's existing streaming interaction abilities. We further incorporate Long-Horizon Interaction Memory (LHIM) to maintain contextual continuity over extended video streams. Evaluations on public streaming benchmarks show that AURA 2.0 achieves leading performance among open-source streaming models, while targeted evaluations further validate its new interaction capabilities. We open-source the model weights and inference code.

A fixed structural edit can preserve a spike sequence under one input and change it under another. We study this dependence using statistics-preserving edge swaps, controlled inputs and finite-horizon parameter certificates. The original 125-chunk numerical model is tested on ten artificial graphs from three connection rules. Despite identical total external drive at every slot, global and focal inputs change the visibility of 4--13 of nineteen fixed swaps per graph. A paired homogeneous and heterogeneous leaky integrate-and-fire validation in brian2-atlas reuses three originals and extends observation from 30 to 300 ms under a constant background drive sufficient for isolated-cell firing. Heterogeneous short-window switching persists, but by 300 ms every tested swap differs from its reference, saturating the binary visibility readout. Fixed-support weight families provide a complementary regional guarantee. Positive certified intervals cover 4,308 of 4,320 LIF direction conditions and include distinct floating-point configurations. A post hoc comparison finds a long-window failure inside the short-window certified interval in 710 of 720 heterogeneous conditions; 394 of 829 direction pairs ordered under both matched-drive inputs change order at 300 ms. Input-specific filtering and threshold-margin scores retain useful ordering but degrade with horizon and do not consistently beat activity controls. Earlier temporal-error, paired-dose and unsuccessful structural-selection results delimit what these findings imply. Together, the evidence links input-conditioned structural effects to finite-horizon fidelity regions in two deterministic kernels, without establishing universal topology rankings, model compression or asymptotic robustness.

We introduce Sage-Taiyi, a metacognition-inspired architecture that uses accumulated scientific understanding to guide what to investigate next and how to investigate it. Through explicit monitoring and control, the system revises its understanding of evidence, assumptions, and unresolved questions, then uses these assessments to adapt research targets and select reusable reasoning methods. On Taiyi-Bench, an AI co-scientist ideation benchmark comprising 265 tasks from 50 papers across ten disciplines, Sage-Taiyi ranks first on all eight metrics of historical rediscovery and open-ended ideation, achieving an Overall score of 54.0 compared with 47.9 for the strongest baseline, AI Scientist-v2. It also performs well on two external scientific ideation benchmarks. Five expert-guided collaborations further connect its proposals to experimental and computational validation across chemistry, materials science, and engineering. In chemistry, reaction-condition optimization raises the yield of cobalt-catalyzed carbonylation from trace amounts to 84% and photocatalytic allylic amination from 13% to 69%. These results support metacognitive regulation as an approach to scientific ideation and its progression toward practical validation.

Fundamental tasks in linear systems analysis and control require evaluating infinite-horizon output energies for autonomous systems and optimal value functions for controlled systems. These quantities are characterized by Lyapunov and algebraic Riccati equations, respectively. In this work, we study promise problems for their evaluation when the system matrices are sparse and represented succinctly by Boolean circuits specifying their nonzero entries, with input length polynomial in the logarithm of the system dimension. For dissipative autonomous systems with coordinate-vector initial states, we define the succinct Lyapunov value problem (SLV). Its extension to linear-quadratic optimal control is the succinct LQR value problem (SLQRV). We prove that SLV is $\BQP$-hard: every problem in bounded-error quantum polynomial time ($\BQP$) reduces to it in polynomial time. We then establish that the more general SLQRV problem admits a quantum algorithm running in time polynomial in the succinct input length. Since SLV is a special case of SLQRV, both problems are $\BQP$-complete. Consequently, unless bounded-error probabilistic polynomial time coincides with $\BQP$, no classical randomized algorithm can solve either problem on all promised instances in time polynomial in the succinct input length.