Cognaptus Insights

TL;DR — L‑MARS replaces single‑pass RAG with a judge‑in‑the‑loop multi‑agent workflow that iteratively searches, checks sufficiency (jurisdiction, date, authority), and only then answers. On a 200‑question LegalSearchQA benchmark of current‑year questions, it reports major gains vs. pure LLMs, at the cost of latency. For regulated industries, the architecture—not just the model—does the heavy lifting. What’s actually new here Most legal QA failures aren’t from weak language skills—they’re from missing or outdated authority. L‑MARS tackles this with three design commitments: ...

Generative models leak. Not because engineers are careless, but because web-scale corpora hide rare, high-influence shards—snippets so unique that gradient descent can’t help but memorize them. A new data-first method, Generative Data Cartography (GenDataCarto), gives teams a way to see those shards in training dynamics and intervene—surgically, not bluntly—before they become liabilities. The one-slide idea Track two numbers for every pretraining sample: Difficulty (dᵢ): early-epoch average loss—how hard it was to learn initially. Memorization (mᵢ): fraction of epochs with forget events (loss falls below a threshold, then pops back above)—how often the model “refits” the same sample. Plot (dᵢ, mᵢ), set percentile thresholds, and you get a four-quadrant map that tells you what to up-sample, down-weight, or drop to reduce leakage with minimal perplexity cost. ...

TL;DR Pair a classic mixed‑integer inventory redistribution model with an LLM-driven context layer and you get explainable optimization: the math still finds near‑optimal transfers, while the LLM translates them into role‑aware narratives, KPIs, and visuals. The result is faster buy‑in, fewer “why this plan?” debates, and tighter execution. Why this paper matters for operators Most planners don’t read constraint matrices. They read stockout risks, truck rolls, and WOS. The study demonstrates a working system where: ...

TL;DR A new paper shows how to insert a sparse, interpretable layer into an LLM to expose plain‑English concepts (e.g., sentiment, risk, timing) and steer them like dials without retraining. In finance news prediction, these interpretable features outperform final‑layer embeddings and reveal that sentiment, market/technical cues, and timing drive most short‑horizon alpha. Steering also debiases optimism, lifting Sharpe by nudging the model negative on sentiment. Why this matters (and what’s new) Finance teams have loved LLMs’ throughput but hated their opacity. This paper demonstrates a lightweight path to transparent performance: ...

The short of it A new study on SWE-agent working over SWE-bench Verified finds that masking old observations (keeping recent turns verbatim, replacing older tool outputs with a placeholder) often matches or slightly beats prompt-based LLM summarization—and at roughly half the cost. The paper also surfaces a subtle failure mode: summaries can elongate trajectories, encouraging agents to “keep going” when they should stop, diluting efficiency and, at times, performance. Why this matters for builders Most production SE agents (debuggers, PR autoresponders, test fixers) rack up spend on two things: tokens and time. Tool logs dominate both. In practice, observation tokens comprise the bulk of an agent’s turn, so trimming them intelligently is the highest‑leverage knob. The results show you might not need fancy, model‑authored summaries; a rolling “mask” window can land on the most efficient frontier—equal or better solve rate, far lower cost—across Qwen3‑Coder 480B, Qwen3‑32B (thinking/non‑thinking), and Gemini 2.5 Flash (thinking/non‑thinking). ...

Thesis: When the job is to read text, reason carefully, and return a precise number (not just a label), ordinary regression heads and vanilla prompting often fail in opposite ways. The paper introduces MENTAT, a lightweight recipe that marries batch‑reflective prompt evolution with a small MLP aggregator over multiple LLM rollouts. The result: tighter calibration and better ranking on tasks where each example demands real reasoning, not surface features. What counts as “Reasoning‑Intensive Regression” (RiR)? RiR tasks look like this: the model must (1) think through the input with step‑wise analysis, and then (2) score it on a real‑valued scale. The paper frames three such tasks: ...

TL;DR — A new framework (EconAgentic) models DePIN growth stages, token/agent interactions, and macro goals (efficiency, inclusion, stability). Its key finding: more patient LLM agents (i.e., slower to exit) can increase inclusion and stability with little efficiency penalty. Sensible—but only if token price formation, data integrity, and geospatial participation are measured rigorously. Why this paper matters DePIN (Decentralized Physical Infrastructure Networks) turns physical capacity—wireless hotspots, sensors, compute, even energy—into token‑incentivized networks. The promise is Uber/Airbnb’s distribution without the platform as rent‑extractor. EconAgentic contributes a general model that: ...

TL;DR Traditional QA treats software as deterministic; LLM apps aren’t. This paper proposes a three‑layer view (System Shell → Prompt Orchestration → LLM Inference) and argues for a collaborative testing strategy: retain classical testing where it still fits, translate assertions into semantic checks, integrate AI‑safety style probes, and extend QA into runtime. The kicker is AICL, a compact agent‑interaction protocol that bakes in observability, context isolation, and deterministic replay. Why this matters for operators and product teams LLM products now look like systems—not prompts. They combine RAG, tools, stateful multi‑turn workflows, and sometimes multi‑agent handoffs. The result is probabilistic behavior plus cross‑layer failure modes. If you keep writing boolean, exact‑match tests, you’ll ship brittle releases and discover regressions in production. The fix isn’t to abandon testing; it’s to move from asserting single outputs to observing semantic behavior distributions. ...

When multi‑turn chats stretch past a dozen turns, users lose the thread: which requests are satisfied, which are ignored, and which have drifted? OnGoal (UIST’25) proposes a pragmatic fix: treat goals as first‑class objects in the chat UI, then visualize how well each model response addresses them over time. It’s less “chat history” and more goal telemetry. What OnGoal actually builds OnGoal augments a familiar linear chat with three concrete layers of structure: ...

TL;DR Neuro‑symbolic architecture (LLMs + Prolog) turns tax calculation from vibes to verifiable logic. The paper we analyze shows that adding a symbolic solver, selective refusal, and exemplar‑guided parsing can lower the break‑even cost of an AI tax assistant to a fraction of average U.S. filing costs. Even more interesting: chat‑tuned models often beat reasoning‑tuned models at few‑shot translation into logic — a counterintuitive result with big product implications. Why this matters for operators (not just researchers) Most back‑office finance work is a chain of (1) rules lookup, (2) calculations, and (3) audit trails. Generic LLMs are great at (1), decent at (2), and historically bad at (3). This work shows a practical path to auditable automation: translate rules and facts into Prolog, compute with a trusted engine, and price the risk of being wrong directly into your product economics. ...