LodeHQ
AI news for people who shipFree, by email

AI · Monday, 9 November 2026 · Issue 165 · 13 stories

Contract AI guards reject over 90% of benign documents

Legal teams cannot use current open-weight prompt-injection guards as simple filters: they flagged more than nine in ten benign contracts and missed every unauthorized-scope claim tested. A cascade using a frontier model only for uncertain cases cut cost 6.5x and nearly matched it.

Models & Releases

Viggle Remake turns whole videos into editable AI scenes

Viggle Remake lets you recast people, swap outfits, or change objects, weather, lighting, and settings across a full video, not just a clip. Give an agent a video and one sentence, and it finds cuts, paints keyframes, renders each window, checks the result, and returns the film with its motion, camera, timing, and soundtrack intact. The app needs no GPU; local runs range from 12 GB cards to 96 GB or more, while the full repository is about 210 GB.

huggingface.co ↗7 min read at source

LiquidAI’s 1.73B model brings long-context agents to 1GB devices

LFM3-1.7B gives developers a 1.73B-parameter text model for local agents and specialized apps, with a 256K-token context window and about 1GB memory use at 4-bit precision. Its tri-hybrid design combines short convolutions with sliding-window and global attention, while quick single-GPU fine-tuning supports domain-specific tools and formats. LiquidAI says its small size makes it unsuitable for agentic coding or complex multi-step reasoning.

huggingface.co ↗9 min read at source

Qwen3.5-4B-3D-Coder runs locally through llama.cpp and Ollama

You can run ccore’s Qwen3.5-4B-3D-Coder locally instead of routing requests through a hosted service. The Hugging Face page provides ready-to-use commands for llama.cpp, Ollama, Docker Model Runner, LM Studio, Jan, and other local apps, including the Q4_K_M variant; llama.cpp can expose it through an OpenAI-compatible server or run it directly in a terminal.

huggingface.co ↗9 min read at source

Research

Contract guards reject more than nine in ten benign documents

Legal-document AI teams cannot treat prompt-injection guards as reliable filters: eleven open-weights guards produced document-level false-positive rates above nine in ten across two contract corpora. The study also finds every guard scored zero on unauthorized-scope claims, while a cascaded setup that checks mid-tier model positives with a frontier model cost six-and-a-half times less than using the frontier model alone and approached its performance.

arxiv.org ↗1 min read at source

ChainFlow gives speculative decoding a 39% lift over EAGLE-3

ChainFlow makes LLM serving faster by using a one- or two-step drafter that predicts through compressed hidden states, rather than generating draft tokens one at a time or denoising from noise. On Qwen3.5 models in vLLM, it beat EAGLE-3 by 39% and DFlash by 11% when drafter size and training data were matched. The paper also releases its code and drafter models.

arxiv.org ↗1 min read at source

Retries can keep an overloaded LLM engine from recovering

For teams serving LLMs through clients or proxies, cancelling timed-out work is not enough: with retries enabled, vLLM restored service in none of the overload runs, delivering only 12% to 38% of calls that arrived afterward. The engine often began retry work less than half a second before timeout, then computed 79% to 85% of the prompt before aborting it. A retry budget or request coalescing worked better, though cancellation could still lose calls.

arxiv.org ↗2 min read at source

Multi-agent debate makes LLM fact-checking attacks harder to defend

A new attack framework, Debate-to-Deceive, generates persuasive false passages that can manipulate LLM fact-checking judgments, raising the risk for systems that rely on them. It pits agents against each other to evolve Logos, Pathos, and Ethos strategies, then has an LLM write the final passage from the winning strategy and debate context. The authors report that it outperforms existing attacks across FEVER, HealthVer, CLIMATE-FEVER, and SciFact while evading most defenses.

arxiv.org ↗1 min read at source

IWIBench finds model editors can rewrite weights without gradients

IWIBench gives researchers a way to test whether one model can directly change another model’s weights, without gradients, while preserving its other abilities. An editor uses four tools to read and write bounded parts of a smaller target model, then a validator checks the patched model against the original across ten text-editing tasks. Results vary sharply: alias acquisition and recoding were easiest, while structured response resisted every attempt.

arxiv.org ↗2 min read at source

Two shell-debugging runs beat mixing in an interactive debugger

For coding-agent builders, adding an interactive debugger is not automatically better: across both GPT-OSS-120B and Nemotron-3-Ultra-550B, two bash-based runs beat one bash run paired with one debugger run on SWE-bench Verified. The study held the agent scaffold fixed across all 500 instances, then introduced pass_X@k, a metric that estimates the chance that at least one of k agents succeeds without confusing tool diversity with random run-to-run variation. The result suggests repeated shell-based attempts may be a better default than debugger access alone.

arxiv.org ↗2 min read at source

Palette makes self-hosted decks match your design language

Palette is a 20B open-weight model for teams that need presentation decks to follow their own design system while keeping confidential content on their infrastructure. It first turns a deck plan into a structured brief, then generates slide-by-slide pptxgenjs code, separating editorial choices from layout work. Trained on synthetic data following IBM Design Language, it ranked first on two of three measured PresentBench domains.

huggingface.co ↗9 min read at source

Identical LLMs can turn shared preferences into a noisy secret key

Two copies of the same closed model can produce matching bits from ordinary preference questions, giving their users a possible shared key over a public, one-way channel without exposing the weights. The experiment asked four models 150 random either-or questions, repeating each eight times, then measured how often model pairs agreed; identical models agreed more than different ones. It is a fingerprinting and coordination result, not yet a running protocol.

matthewritch.com ↗1 min read at source

Products & Industry

Measure coding agents by merged software, not tokens consumed

Teams using coding agents should track cost per merged pull request, not token usage, because one agentic task can consume 10 to 50 times more tokens than a basic chat prompt. The measure divides total AI-assisted development cost by reviewed, merged changes, then checks rework and defects so cheap but poor code does not look efficient. The article also recommends routing hard planning to frontier models and routine coding or testing to cheaper open-weight models.

technologyreview.com ↗5 min read at source

Tools & Open Source

Transformers shrinks model tests enough to catch bugs before merge

Transformers can now put lightweight, real-model integration checks on every pull request, giving maintainers a way to catch regressions before they reach main without keeping GPUs busy all day. Its proposed Model Integration Contract packages a shrunken model fixture, a runnable usage example, and reviewed output baselines; the UMT5 fixture used 66 MB of weights versus 1,227 MB for the source checkpoint, and caught a decoder bug that let it see future tokens during training.

huggingface.co ↗7 min read at source

You’re caught up.

That was all 13 stories in the 9 November issue.