Recurrent context compression controls context growth in long-horizon agents, but its behavioral effects remain poorly understood. In this preliminary empirical study, we show that compression can weaken the influence of recent interactions, increasing blocked actions, repeated exploration, and instability across runs. Motivated by these observations, we introduce TRACE, a verifier-guided framework that evaluates individual compaction events through paired closed-loop continuations from the same environment state and uses summary preferences to optimize a natural-language compression prompt while keeping all models frozen. Initial results on AppWorld show improvements over existing compression baselines in task performance, multi-run reliability, and context–execution efficiency. These findings provide early evidence for boundary-local evaluation as a promising direction for reliable agent context compression.
Natural generation allows Language Models (LMs) to produce free-form responses with rich reasoning, but the lack of guaranteed structure makes outputs difficult to parse or verify. Structured generation, or constrained decoding, addresses this drawback by producing content in standardized formats such as JSON, ensuring consistency and guaranteed-parsable outputs, but it can inadvertently restrict the model's reasoning capabilities. In this work, we propose a simple approach that combines the advantages of both natural and structured generation. By allowing LLMs to reason freely until specific trigger tokens are generated, and then switching to structured generation, our method preserves the expressive power of natural language reasoning while ensuring the reliability of structured outputs. We further evaluate our approach on several datasets, covering both classification and reasoning tasks, to demonstrate its effectiveness, achieving a substantial gain of up to 27
Multi-Scalar Multiplication is a critical operation in most pairing-based zero-knowledge proofs. In a lot of studies, memory limitations have often been reported to be the primary bottleneck preventing the calculation of larger MSMs. In this paper, we are particularly interested in the acceleration of this operation on devices with limited memory. Pippenger’s algorithm (also known as bucket method) is the most efficient and, consequently, the most widely used method to calculate Multi-Scalar Multiplications. We propose an optimization of Pippenger’s algorithm which is at least as efficient as the original, and significantly more effective when operating under limited memory. The main idea is to use an adapted number of buckets depending on the available memory instead of 2^w - 1 . We conducted tests on the curve BLS12-381 with Multi-Scalar Multiplications ranging from 2^8 to 2^14 points. The results obtained demonstrate that we have a very significant gain (up to 40% ) for very limited memories. This gain gradually decreases as more memory becomes available, until we achieve performance comparable to Pippenger’s once memory is no longer limited. For example, in a Multi-Scalar Multiplication with 2^13 points, we observe a gain of 40% with only 1 KB of memory, 20% with 15 KB, 15% with 35 KB, and so on, down to be equivalent to Pippenger’s algorithm once memory is no longer a constraint.
As lifelong learning agents accumulate lifelong growing skill banks, retrieving the correct skill becomes an increasingly important bottleneck. Most current skill retrieval methods treat each skill as one flat document by concatenating fields such as the name, description, and body. However, skills are naturally structured, multi-field objects, where each field provides different information about when and how the skill should be used. In this work, we study whether preserving this structure improves skill retrieval. We represent each skill as its separate components, and compute sparse and dense similarities for each field independently, exposing a naturally tensorized, field-aware representation of the skill bank. We then combine these field-level scores either with uniform weights or with a small learned MLP. Across two different skill retrieval benchmarks, SkillRet and SRA-Bench, we find that keeping fields separate improves hybrid retrieval, and learning over the field-level scores gives the strongest and most consistent results. Our field-aware MLP reaches 77.95 Recall@10 on SkillRet and 83.78 Recall@10 on SRA-Bench, outperforming the corresponding concatenated learned baselines. We also find that the advantage grows as the skill bank becomes larger, suggesting that field-aware skill retrieval becomes especially useful in the setting where retrieval is most difficult. Our results show that skill representation itself matters, and that simply preserving the structure already present in skill files can substantially improve retrieval.
We present a deep neural network approach for encoding microphone array signals into Ambisonics that generalizes to arbitrary microphone array configurations with fixed microphone count but varying locations and frequency-dependent directional characteristics. Unlike previous methods that rely only on array geometry as metadata, our approach uses directional array transfer functions, enabling accurate characterization of real-world arrays. The proposed architecture employs separate encoders for audio and directional responses, combining them through cross-attention mechanisms to generate array-independent spatial audio representations. We evaluate the method on simulated data in two settings: a mobile phone with complex body scattering, and a free-field condition, both with varying numbers of sound sources in reverberant environments. Evaluations demonstrate that our approach outperforms both conventional digital signal processing-based methods and existing deep neural network solutions. Furthermore, using array transfer functions instead of geometry as metadata input improves accuracy on realistic arrays.
