iiinova — where your digital self keeps evolving

Nir Grinberg

@nirg.bsky.social

185 followers 174 following 28 posts

Assistant prof. at BGU in the field of Computational Social Science.

Nir Grinberg @nirg.bsky.social · 07/05/2026

5/8 📊 The results? CoLA consistently outperforms all baselines. RAG captures surface-level style (like POS distributions), but fails in reasoning. CoLA-generated reasoning is so faithful that an LLM trained specifically on each judge cannot tell the difference from the actual judge. 🤯

Personalization performance across BLEU, RougeL, BertScoreF1, and POS-JSD.

Model performance in the Author Discernibility Task.

100

Nir Grinberg @nirg.bsky.social · 07/05/2026

4/8 🧠 For training, we used Chain-of-LoRA (CoLA) in two steps: 1️⃣ Causal LM tuning (adapts to general writing style). 2️⃣ Instruction tuning (focuses on legal reasoning). All using a single GPU and Gemma-4B model, which was super fast.

The ablation study showing performance as a function of training size and LoRA rank

100

Nir Grinberg @nirg.bsky.social · 07/05/2026

🚨 Excited to share our new paper to appear in the Findings of ACL 2026! "JudgeMeNot" asks: Can LLMs learn the specific reasoning signature of an individual judge? We tackle this in the low-resource setting of Hebrew case law. 🔗 arxiv.org/abs/2604.18041 🧵👇

The illustrative diagram shows our approach for training personalized LLMs to emulate Judge A (textualist) vs. Judge B (purposivist) and how those trained models might respond to a legal query.

152

Nir Grinberg @nirg.bsky.social · 13/01/2026

Key finding 3: This also works at the state and city (!) level, including "holdout cities" where official UI numbers are sparse or irregularly updated. As expected, accuracy scales with platform penetration and unemployment shocks. 7/N

110

Nir Grinberg @nirg.bsky.social · 13/01/2026

Key finding 2: Our approach consistently outperforms industry consensus forecasts and can improve predictions of US UI claims up to two weeks ahead of official releases. That’s two weeks of additional lead time for policymakers. 6/N

110

Nir Grinberg @nirg.bsky.social · 13/01/2026

Key finding 1: Capturing linguistic diversity matters. Training LLMs with active learning lets us detect many more ways people talk about job loss, producing a far more representative sample of unemployed users than existing approaches. 5/N

110

Nir Grinberg @nirg.bsky.social · 13/01/2026

Why this matters: In March 2020, weekly unemployment insurance claims jumped from 278K to nearly 6 million in two weeks. As official data lagged, policymakers were flying blind about where the shock was hitting and who was being affected. 2/N

100

Nir Grinberg @nirg.bsky.social · 05/12/2023

Finally, looking at the demographic composition of consumption "types", we find that the media-oriented clusters (exc. superconsumers) have older individuals, more women, and more registered Democrats. 5/

Sociodemographic characteristics among different political exposure types. Sample averages are marked in a gray dashed line. Ninety-five percent bootstrapped CIs are shown (mostly occluded due to their small size). CI = confidence interval.

121

Nir Grinberg @nirg.bsky.social · 05/12/2023

Americans also vary in the breakdown of actors that populate their feeds, but interestingly, the bulk of the population gets half or more of their political exposure from *traditional sources*—media organizations, journalists, and politicians. 3/

The composition of political exposure across clusters. The share of politics curated by different actor types (y-axis) across clusters (x-axis). Darker-colored bars represent direct exposure to media organizations, journalists, politicians, OLs, and social peers. Lighter-colored bars represent indirect exposure to media organizations, journalists, politicians, or opinion leaders through social peers. OL = opinion leader.

121

Nir Grinberg @nirg.bsky.social · 05/12/2023

People's political feeds mostly map onto 8 distinct types that vary in the amount of politics they get, both in absolute #'s and as % the feed as a whole. Still, for nearly 90% of the population, about 1 in 12 posts from their network are political. Quite an engaged public! 2/

Prototypical types of individual political exposure. Each point in panel (A) represents the political exposure of a single panel member, reduced to two dimensions using the UMAP algorithm, and colored by the cluster assignment obtained from HDBSCAN. Panel (B) shows the median number of political tweets available to individuals per day (left bars), and their percentage out of all tweets available to them on Twitter (right bars). Cluster labels and their share in the population are specified on the x-axis. Colors are consistent between the two figure panels. Ninety-five percent bootstrapped CIs are omitted from the figure due to their small magnitude, which are upper bounded by twenty-seven exposures to tweets and 0.28 percent, respectively. OL = opinion leader; CI = confidence interval; UMAP = Uniform Manifold Approximation and Projection.

131

Nir Grinberg @nirg.bsky.social · 05/12/2023

🚨New paper🚨 out in the International Journal of Press/Politics w/ Assaf Shamir and @jenny-oser.bsky.social 🎉 Here's what we learned from studying the composition of political content available to 600k+ registered U.S. voters on Twitter during the 2020 election. doi.org/10.1177/1940... 🧵👇

12011