One-Line Summary

A compact vision-language model pairs counting with count-accurate generation, while new methods cut long-context memory and add physics signals to forecasting.

Research Papers

ABACUS bridges counting and image generation in one model

ABACUS is a single vision-language system that can answer “how many” questions in images (objects, crowds, and referring expressions) and also generate images that obey requested counts, without training for each specific benchmark. It adapts a 3‑billion‑parameter foundation model to object localization and to “count‑faithful” generation in the same framework. ¹

To localize and count reliably, ABACUS adds density‑aware adaptive zooming guided by objectness maps, so the model looks closer where items cluster. It further trains a boundary‑aware counting policy to avoid crop‑edge mistakes, and uses a cycle‑consistent self‑critique where the understanding branch checks whether generated images respect the requested counts—closing the gap between perception and generation without external annotations. ¹

On seven benchmarks, the authors report state‑of‑the‑art results, outperforming both task‑specific counters and larger generalist models. The headline claim is that unifying counting and generation reduces task fragmentation while improving accuracy, all while staying on a relatively small 3B backbone. ¹

Why it matters: one system that both understands and creates count‑accurate scenes can simplify data augmentation, simulation, and QA pipelines in domains like retail inventory, crowd analytics, or ecology. Watch how robust the “count‑faithful” promise remains outside curated datasets and whether third‑party implementations reproduce the reported gains. ¹

InfoKV compresses long-context memory with uncertainty signals

When large language models (LLMs) process long prompts, the key–value (KV) cache that stores token representations can balloon; InfoKV trims it by selecting tokens using information‑theoretic cues, not just attention weights. The method introduces Forward Influence—a forward‑looking metric that measures how keeping a token changes future contexts—and blends token‑level predictive uncertainty with layer‑wise representation changes to decide what to keep. ²

Across long‑context reasoning evaluations with Llama‑3.1, Llama‑3.2, and DeepSeek‑R1, InfoKV consistently outperforms attention‑only compression in both long prefilling and decoding, indicating a path to lower memory use without sacrificing accuracy. For practitioners, this implies more headroom for longer inputs or larger models within the same hardware constraints. ²

EO-WM uses physics-aware signals to forecast Earth observation

EO‑WM reframes Earth Observation (EO) forecasting as a weather‑driven world‑modeling task and builds a video diffusion transformer that explicitly conditions on physics‑informed signals: a climatological baseline, weather anomalies, and cumulative stress (e.g., heat/drought) over time. The paper also proposes two diagnostics—the Extreme Summer Benchmark for severity‑aware vegetation degradation and a Seasonal Matched‑Pair Benchmark to test whether forecasts respond correctly when weather forcing changes. ³

In experiments, EO‑WM reduces error in predicted Normalized Difference Vegetation Index (NDVI) decline amplitude by a relative 5.63% and improves directional hit rate by a relative 7.80%, while remaining competitive on pixel‑level metrics. The authors state they plan to open‑source the benchmarks and model, which would help others test weather‑response fidelity directly. ³

Sequence probability doesn’t always mean a more correct answer

This study asks a simple question with big implications: when does making a continuation more probable under the model actually make it more correct? The authors measure the link between sequence probability and correctness across decoding methods, hyperparameters, datasets, and repeated answers to the same prompt. ⁴

They find higher sequence probability often correlates with correctness across different prompt‑answer pairs in a fixed dataset, but this does not generally carry over to decoding choices: cranking hyperparameters or switching methods to increase probability doesn’t reliably boost accuracy, and probability poorly predicts correctness across repeated samples for the same prompt. Practical takeaway: treat probability as a coarse signal—use self‑consistency or verifiers carefully, but don’t assume “more likely” equals “more right.” ⁴

Why It Matters

Today’s papers point to a pragmatic trend: combine capabilities in smaller models (counting plus generation), squeeze memory smartly (information‑aware KV pruning), and bake domain physics into forecasts. At the same time, decoding work reminds us that simple probability‑maximizing tweaks are not a shortcut to correctness—evaluation and verification still matter.