AI FRONTIER
Signals worth understanding before they become obvious.
One Model to Loop Them All: Elastic Latent Recurrence for Language Models
The paper presents Elastic Latent Recurrence (ELR), a recipe for training looped language models so that at one loop they cost the same to serve as a non-recurrent model of the same size and reach similar accuracy, with each additional loop buying more accuracy for more inference compute.
- On 4B dense and 30B-A3B MoE hybrid models, ELR at one loop is 0.52 points below and 1.04 points above its matched baseline on a ten-task average, and four loops add 7.25 and 4.75 points.
- In MoE models, reusing first-loop expert routes on later loops cuts mixed-depth update time by 11.7% with no measurable change in accuracy.
- At inference, keeping only the last loop's attention KV cache holds attention KV memory at its one-loop level while retaining most of the gain from additional loops.
- Carried through post-training, the recipe yields Koi-4B, a single checkpoint that scores an estimated 18 on the Artificial Analysis Intelligence Index at one loop, matching Qwen3.5-4B, and 29 at four loops.
Because one checkpoint can trade loop count for latency and accuracy, looped models become viable for latency-sensitive serving rather than requiring a separate non-recurrent model.
SpireBench: Evaluating Long-Horizon Strategic Play of Agentic Models with Slay the Spire
The paper introduces SpireBench, in which models play the complete commercial deck-builder Slay the Spire at its highest difficulty to evaluate long-horizon strategic reasoning across 20 large language models.
- The paper evaluates 20 large language models at Slay the Spire's highest difficulty.
- The paper reports that frontier models beat heuristic baselines and match the average human player but remain far behind experts.
- The paper reports that the strongest models win the first act as often as experts, then fall behind as the run lengthens.
- The paper reports that a comprehensive, self-consistent renaming of every game term shows part of performance rests on prior knowledge of the game, and models do not improve from the memories they write across runs.
For agent builders, SpireBench offers a long-horizon benchmark that locates where agents fail and points to specific weaknesses in risk judgment and memory use.
Beyond Binary Verifiable Rewards: Behavioral Self-Rewarding for SWE Agents
The paper introduces Behavioral Self-Rewarding (BSR), a reward-refinement method that reuses the policy to compare observable debugging behavior among trajectories from the same task with the same verifier outcome, aggregating pairwise preferences into task-local scores and converting them into bounded auxiliary reward...
- BSR reports consistent improvements on SWE-bench Verified and SWE-bench-Live across the evaluated Qwen3-14B and Qwen3-32B configurations.
- The method reuses the policy to compare observable debugging behavior among trajectories from the same task with the same verifier outcome, without a separately trained reward model.
- BSR aggregates pairwise preferences into task-local scores and converts them into bounded auxiliary rewards, introducing reward contrast while preserving the verifier-defined pass/fail ordering.
- BSR provides additional supervision from already-collected trajectories without requiring new environment interactions.
For teams building SWE agents, BSR indicates that the policy's own comparative judgments can serve as a complementary learning signal beyond binary test pass/fail, without new environment interactions or a separate reward model.
Kairos: Toward a Byte-Level Foundation Model for Every Modality
Kairos combines bidirectional gated-DeltaNet linear attention with sliding-window attention over shared projections (LiZAttention2), block-wise Attention-Residual aggregation, a sparse Mixture-of-Experts feed-forward layer and a multi-scale byte-level codec, trained with a uniform-noise denoising objective over one byt...
- In overfitting tests on a single 440-byte sentence at 1–15M parameters, text only, LiZAttention2 and window attention alone reached a minimum loss near 10^-3 at 3,000 steps.
- After 600 steps, mean minimum loss ranged from 0.018 (patch codec) to 0.354 (depthwise convolutional codec), with identical ranking on three seeds.
- The Discussion sizes a 200M-total/25M-active target analytically, but none of its throughput figures is measured; code and a Python package are available.
For builders, Kairos is currently a reproducible convergence sanity check, and its byte-level multimodal design still needs the planned pretraining run to show generalization or multimodal value.
ForcingBooster: Adaptive Patch-wise Denoising for Fast Autoregressive Video Diffusion
The paper introduces ForcingBooster, a framework that dynamically allocates denoising steps to individual video patches, stopping easy patches early while giving difficult patches more refinement steps. It reports reducing the total number of denoising steps applied across video patches by about 50% while maintaining v...
- The paper reports reducing the total number of denoising steps applied across video patches by about 50%.
- The paper reports visual quality comparable to the four-step baseline.
- The method is compatible with most existing few-step autoregressive diffusion models and requires only a lightweight routing module with modest training.
- The paper was submitted on 8 Oct 2026 as arXiv:2610.12625 by Zekai Gu, Peng Wang, Jiahao Lu, Yongsen Mao, Cheng Lin, and Yuan Liu.
For teams building real-time interactive video generation and video world models, patch-wise denoising budgets can shift computation from easy regions to difficult ones without changing the few-step autoregressive diffusion backbone.
Google DeepMind releases EmbeddingGemma 2, a 740M open multimodal embedding model
EmbeddingGemma 2 maps text (including code), images, video, and audio into a single unified 768-dimensional vector space under an Apache 2.0 license, designed to run on consumer hardware such as mobile devices and laptops.
- 740M total parameters: 270M text backbone (130M transformer + 140M embedder), 170M vision encoder, and 300M audio encoder.
- The paper reports MTEB multilingual v2 at 61.36 and MTEB code v1 at 78.68, a roughly 14% improvement over EmbeddingGemma 1's 68.76 on code.
- Matryoshka Representation Learning supports truncation to 128d, 256d, 512d, and 768d, enabling up to a 6x reduction in vector storage costs.
- 8,192-token context window, 100+ languages supported, Apache 2.0 license.
Builders can serve text, image, video, and audio retrieval from one 768-dimensional embedding space on consumer hardware, and selectively load encoders to cut the effective footprint from 740M to 270M.
MagicBeanStalk: Escaping the RLVR Zero Advantage Trap with Recursive Data Synthesis
The paper introduces MagicBeanStalk, a top-down decomposition and bottom-up recursive data-synthesis algorithm that grows seed problems into policy- and budget-calibrated training problems while preserving the seeds' reasoning blueprints.
- The paper reports that under a fixed target policy and rollout budget, roughly 88% of a large open-source math corpus pooled from widely used RLVR datasets is solved in every rollout or failed in every rollout.
- The paper reports that across 8 math benchmarks, continual RLVR on MagicBeanStalk problems improves gpt-oss-120b from 61.2 to 69.8 (+8.6) and DeepSeek v4Flash from 72.4 to 77.9 (+5.4) on a 16K-token budget.
- The paper reports that training on the full 78K source corpus improves by only +1.1, and its best filtered real data subset by +5.8.
- The paper reports that in finance it synthesizes over 14K verifiable problems from only 340 solutions and improves over base by 6.9 and 2.4 points on SecQue and FinanceBench, respectively.
For RLVR data builders, the useful signal is whether a problem yields non-zero advantage under the target policy and rollout budget, not corpus size, and recursive synthesis can calibrate difficulty while preserving reasoning structure.