01 DAILY RESEARCH SIGNALS / 2026-10-13
A curated stream of quantitative research, open-source projects, and AI engineering advances—screened, ranked, and structured for faster research decisions.
All sources checked. Pipeline operating normally.
02 CURATED INDEX
Browse by content type or search titles and tags.
INDEXED 15
SELECTED SIGNALS 03
This project proposes context window optimization for AI coding agents, claiming that sandboxing tool output cuts context usage by roughly 98%, while persisting session memory and enforcing routing across 17 platforms via MCP and hooks.
rea is a GitHub project proposing to use agents for reverse engineering, from app behavior down to native binaries, shifting manual reverse-engineering workflows toward AI agents and offering a new automation angle.
Vercel Labs has published the open-source project npx skills on GitHub, offering an npx-based approach to managing agent skills. Its novelty lies in packaging agent capabilities as an open tool for easier installation and invocation, which is practically relevant to developers building AI tools.
SELECTED SIGNALS 09
This work introduces SynAct, a framework that allocates expert feedback at the individual decision level rather than at the task level, using mutual information to identify consequential decisions and improve multi-step LLM agent accuracy under scarce expert attention.
This study systematically reveals a unique hidden safety threat in large reasoning models (LRMs): harmful content can be embedded in intermediate Chain-of-Thought (CoT) steps while the final answer appears benign. Existing state-of-the-art moderation tools suffer substantial performance degradation on CoT moderation (F1 drops by up to 35.2%). To address this, the authors release the ReasoningShield benchmark (9.2K annotated pairs) and train lightweight 1B/3B models that achie
This study applies reflective prompt evolution (GEPA) to two system prompts of a multi-agent deep-research agent, evaluating transferability under a topic-disjoint split. Its novelty lies in showing that the validation-minus-held-out gap is small-sample selection rather than memorization, and identifying decomposed per-dimension feedback as the causal ingredient for transferable gains, implying credible evaluation requires held-out topics, multiple seeds, and verifier ablatio
This study builds a reproducible benchmark for the alternating row and column attention patterns unique to tabular in-context learners, comparing Torch SDPA, FlashAttention-2/3/4, vLLM, and SageAttention on three GPUs for forward and backward throughput. It is novel in extending efficient attention evaluation from one-dimensional language sequences to two-dimensional tabular settings, showing that the optimal backend varies with column/row attention, sequence length, head dim
Holding detector configuration fixed and varying only the test boundary, this study quantifies how reuse of Trust-Hub host variants across train and test folds inflates apparent performance. It provides empirical grounding for family-aware holdout evaluation in hardware-Trojan benchmarks.
This study proposes the RTCARE model, which incorporates realized quantile and realized ES constructed from intraday high-frequency returns into the dynamics and measurement equations of conditional expectiles for joint VaR and ES forecasting. Its novelty lies in combining low-frequency tail-risk forecasting with high-frequency realized tail information; it has potential practical value for quantitative risk management and backtesting.
This study targets transaction narrative text, which credit risk modelling has long overlooked, and proposes an explainable method to convert it into behavioural features. Its novelty lies in using natural language processing and regular expressions to build a categorisation algorithm that maps unstructured transaction descriptions into interpretable spending categories, and in validating their incremental predictive value within a traditional scorecard framework. For quantit
New contribution: in discrete-time incomplete markets, restricting strategies to bounded finite-dimensional parametric classes makes the admissible gain set compact and convex in L². The paper proves that as risk aversion tends to zero, restricted entropic hedging gains converge in L² to the unique restricted quadratic hedge, providing a unified framework for both optimization problems.
A backward actor-critic stochastic policy-gradient scheme is proposed, combining discrete dynamic programming with PPO and one-hidden-layer neural-network approximations of value function and control policy. Continuous controls are represented by a squashed Gaussian policy built on a C-vine correlation-matrix representation that enforces positive definiteness by construction. Beyond robust prices, the trained critic's spatial gradient approximates the superhedging strategy, a
SELECTED SIGNALS 01
This page collects a set of quant trading links newly featured by Quantocracy as of October 12, 2026, and introduces Quant Observer, a curated directory of quant tools, data, and research. Its value lies in providing a centralized entry point to quant resources for readers to access quickly.
SELECTED SIGNALS 02
Using PewDiePie's Odysseus as a case study, the video breaks down how an agent harness runs, how tool calling works, how to run models locally with llama.cpp and Hugging Face, and how to fine-tune a small model like Qwen 3.5 9B on agent traces. It is a practical reference for developers who want to understand or build their own agents.
Token Arcade presents structured output and constrained decoding as an interactive game, showing how a model's top prediction can violate rules and helping users understand next-token probabilities and allowed choices.