68 materials
The study compares two ways of applying ready-made knowledge packages: incorporating their instructions into an agent’s working context or assigning subtasks to separate agents. Delegation is more effective when inputs, outputs, and procedures are clearly defined, but it requires additional coordination overhead.
The article presents XAI-Arena, a system that compares AI explanations across multiple criteria, taking different models and user interests into account. LLM evaluations align closely with human judgments (Spearman’s ρ = 0.693), suggesting that comparative analysis of explanations can be scaled.
The article describes a C++ implementation for a language model with binary spike gating: sparse projections are processed in INT8, and computations account only for active weights. In a single-threaded test, the early INT8 version achieved 23.31 tokens/s versus 9.82 for FP32 and reduced memory usage for weights from 3355.2 to 1087.4 MiB; switching dense projections to INT4 reduced decoding throughput.
The article presents a dual-stream architecture that combines LIF spiking dynamics with continuous context processing and local attention; four models are trained from scratch on Chinese and English data. Encoder probes show around 90% zero spikes, but average perplexity is 7.7% worse than that of a similarly sized GPT-2, and measured generation is noticeably slower—the results outline a trade-off between sparsity, quality, and computational efficiency.
The article presents a mechanism that selects individual generations for retraining based on advantage magnitude and discards overly stale data, supplementing it with fresh examples. Across three Qwen3-Base scales, the method outperformed GRPO and standard replay on math benchmarks, while also improving the balance between accuracy and token usage.
A comparison of efficient-KAN architectures on noisy and small samples shows that at σ=0.1, ChebyKAN's error on a clean reference increases by a factor of 10.6, whereas the proposed ER-KAN's error increases by only a factor of 1.4. Robustness is provided by shared Gaussian RBF bases, gradual noise injection during training, and entropy-weighted regularization; the work also introduces a noise degradation ratio metric and demonstrates a 4.2-fold reduction in MSE relative to MLP on a PINN for a damped oscillator, while honestly reporting the failure with the Burgers equation.
The LOCKS method stores a separate low-rank representation of keys for each context page, so block importance estimation relies only on this summary, while the candidate keys themselves are not read. On long tasks, this approach keeps quality close to a full cache while accessing roughly 2% of tokens and noticeably reducing decoding latency.
The framework groups weak nodes into replica sets and distributes token processing phases differently to efficiently utilize limited resources. In tests, this approach nearly halved average response time compared to the well-known Splitwise scheme.
Researchers propose viewing the Text-to-SQL field as a consolidated leaderboard, where results are distributed by the degree of model autonomy in generating SQL queries. They conduct measurements on Spider and show that successes do not always transfer to other benchmarks, and that autonomy provides robustness but requires significant computational resources.
