AI FRONTIER
Signals worth understanding before they become obvious.
Flint-interactive: A language for creating interactive data visualizations
Microsoft Research released Flint-interactive, a language built on Flint that lets AI agents define clicking, filtering, zooming, highlighting, and cross-chart interactions through semantic events and update operators, with interaction logic handled by the Flint compiler.
- Flint-interactive resolves raw triggers such as clicks, drags, and key presses into semantic events that identify chart elements and data.
- Users specify interactions by mapping semantic interaction events to chart update operators such as set style, set annotation, and set viewport, without manipulating renderer-specific marks.
- It is usable in chat or canvas environments through Flint-MCP and MCP Apps, where an agent can generate an interaction_spec alongside the Flint chart spec.
- Flint Interactive Canvas lets apps process semantic events with event handlers to filter a table, update another chart, or change application state.
For builders of data agents, delegating interaction logic to the compiler can cut the token and time cost of repeatedly generating and debugging interactive charts while reducing layout and interaction failures.
WorldStream: Streaming Interactive 3D Worlds with Hierarchical Planning and Persistent State
The paper presents WorldStream, an open-ended 3D world model that extends playable terrain beyond any pre-specified map, acts on open-ended language in the running world, and restores visited places with their edits. A hierarchical planner over a typed decision model emits structured decision state, and deterministic l...
- Macro fixes settlement, ecology, transport, and hydrological structure across a group of tiles, while Micro specializes that contract into tile semantics under measured terrain and accepted neighbor commitments.
- Streaming separates reach: plans run ahead along the predicted corridor, compilation happens only where movement will soon need it, and geometry behind the user is released.
- One world record retains plans, seeds, object identities and mutable state, boundary commitments, and edits by attribute, revised as the user advances and edits, and returning tiles recompile from it.
For teams building world models, separating planning, compilation, and a persistent state record is a concrete architecture path to tile-to-tile coherence and editability under bounded budgets.
When Thoughts Become Memory: Unifying In-Context Learning and Fine-Tuning
The paper gives a certified rule, built on an exact law for information value, for whether unseen data belong in a pretrained model's context or its weights, and an anytime-valid test makes that choice from held-out queries.
- The paper reports the test picks the better of GPT-2's cache and fine-tune in 61,948 of 64,000 query streams over 32 comparisons.
- The paper reports it errs in at most 3.3% of null streams.
- The paper reports one patch that sees a shape feature cuts an occupancy network's code length there a hundredfold, while eight that miss it nearly double it.
- The paper reports that on a yes-or-no claim the valuation recovers a published evidence adjudicator, with AUROC 0.91.
This certified rule lets builders decide from one measurement and held-out queries whether data should enter context or weights, reducing the cost of choosing between in-context learning and fine-tuning.
OpenRouter new model: Recursive: Flint (free)
Flint is an open-weight mixture-of-experts model from Recursive with about 2.5B active parameters out of 40B total, built by Recursive's autoresearch system Eureka Machine for coding agents.
- 40B total parameters with about 2.5B active parameters
- Context window of 262,144 tokens
- Released on October 21, 2026
- Free endpoint with rate limits; accepts tools for function calling and supports response_format for JSON output without JSON-schema enforcement
For builders of coding agents, Flint offers a zero-cost open-weight option with 2.5B active parameters and a 262K context window to test.
SimHack: Predicting the onset of detectable reward hacking before RL training
The paper introduces SimHack, a simulator that learns reward hacking training dynamics from a handful of smaller-scale reference RL runs and generates counterfactual simulation runs of unseen target recipes without costly retraining, predicting the distribution of the reward hacking onset step.
- The paper reports that SimHack predicts the median onset to a relative error of 11.3% and achieves across-recipe rank correlation of 0.85 on unseen target recipes.
- SimHack generalizes to differences in model scale, curriculum, number of rollouts, and learning rate between reference and target recipes.
- As an application, SimHack screens candidate RL curricula and discovers curricula over the same set of scenarios that train for 96.6% longer without hacking.
- The paper finds that the initial hacking rate alone does not determine onset, which varies substantially across training recipes and repeated runs.
Builders can pre-screen curricula with SimHack before expensive RL runs to choose lower-risk recipes at low cost.
KVLearn: Learning-Based KV Cache Retention for Disaggregated LLM Serving
The paper presents KVLearn, a learning-based framework for cost-aware KV-cache retention in disaggregated LLM serving, combining a Prefix Reuse Predictor (PRP), a Cost-Aware Retention Score (CARS), and an Adaptive Threshold Controller (ATC), evaluated on text and multimodal workloads.
- The paper reports end-to-end time-to-first-token reductions of up to 56% versus No-Cache, 38% versus LRU-Pool, and 33% versus Mooncake-style baselines.
- The paper reports inter-node KV transfer volume cut by up to 53% versus LRU-Pool.
- On MM-Session, throughput remains within about 5% of oracle.
- PRP estimates reuse probability from structural and temporal prefix features without modifying model weights; the paper was accepted to ACM SYSTOR 2026 and code is available.
For teams building disaggregated inference serving, KVLearn shows that folding recomputation, transfer, and storage costs into cache admission decisions can cut both time-to-first-token and inter-node bandwidth without touching model weight...
MPS in PyTorch 2.15: FP8 support and 5–6.5× faster MoE decoding
PyTorch 2.15 brings FP8 tensors, native grouped matmul and dedicated Metal kernels to the MPS backend, making MoE model decoding 5–6.5× faster than 2.14 and requiring macOS 15 and Python 3.11 or later.
- torch.float8_e4m3fn tensors now work on MPS at one byte per value, half the storage of FP16 or BF16, and F.scaled_mm supports tensorwise and rowwise scaling.
- On an M3 Ultra with torch.compile, Qwen3.6-35B-A3B goes from 9.4 to 61.3 tokens/s and LFM2.5-8B-A1B from 23.6 to 123.6 tokens/s.
- Across 96 BF16 layer shapes from real models, connecting GEMV kernels to F.linear gives a 1.23× geometric-mean speedup.
- Wan2.2's VAE decode peak driver memory drops from about 270 GiB to 8 GiB, and FP32 softmax over a [1, 128256] logits row drops from 57 µs to 7 µs.
Mac-based builders can cut memory with FP8 weights and get large LLM decoding gains on both MoE and dense models, provided they move to macOS 15 and Python 3.11.
Big Claims, Small Margins: Do Reported GraphLLM Gains Survive Simple Baselines?
The paper compares four baseline representatives, tuned with validation-only selection, against the highest mean each of three recent studies reports for an LLM-based method on matched splits, and documents evaluation risks in released code.
- Correct & Smooth on TF-IDF features trails the reference by a median of 2.3 percentage points, and an MLP on frozen e5-large-v2 embeddings by 2.8 points.
- Correct & Smooth is within 3 points of the reference or better on 15 of 24 tasks, but the reference is significantly higher on 12 of 22 testable tasks.
- Reports of the same method on the same split differ by up to 12.5 points across papers whose configurations are not documented to match.
Builders of GraphLLM node-classification pipelines should run tuned TF-IDF and frozen-embedding baselines on the same splits and audit data loaders and tuning protocols before trusting reported state-of-the-art gains.