iiinova — where your digital self keeps evolving

We hope that our prediction task can support the development of better value representations. We’re also excited to study interactions values in more realistic alignment post-training setups, which could support the development of ecosystems of models representing a wider range of values.

120

Andy Liu @andyliu.bsky.social · 21h

We label the original Constitutional value set using our taxonomy as well as two existing taxonomies of LLM values. We find that our labelling better recovers generalization than Values in the Wild and LitmusValues, both of which are built on textual descriptions of values.

120

Andy Liu @andyliu.bsky.social · 21h

We also introduce the ValueMap pipeline to cluster value representations, leading to the first taxonomy of LLM values grounded in generalization dynamics. Within a larger value set from Anthropic’s Values in the Wild, we identify four main clusters: attunement, rigor, stewardship, and integrity.

120

Andy Liu @andyliu.bsky.social · 21h

We hypothesize that value representations that perform better at our task can also support the empirical study of alignment targets. To support this, we study multi-value alignment targets, finding that more coherent targets (i.e. whose values have similar representations) are more robust.

130

Andy Liu @andyliu.bsky.social · 21h

We benchmark different value representation methods by how well they predict alignment generalization effects. We find that methods based on model activations when applying values in context (such as persona vectors) significantly outperform methods based on textual descriptions of the values.

120

Andy Liu @andyliu.bsky.social · 21h

We extract 66 values from Claude’s Constitution, and compute a generalization matrix, where each row represents a value that we fine-tuned on, and each column represents a value that we evaluate the fine-tuned checkpoint on. Matrices are similar across model families, sizes, and fine-tuning methods.

130

Andy Liu @andyliu.bsky.social · 21h

Our work introduces an alignment generalization prediction task. Given a set of values, we fine-tune single-value checkpoints on each value in our set, then evaluate how this changes each individual checkpoint’s adherence toward all other values in the set.

120

Andy Liu @andyliu.bsky.social · 21h

LLM developers train their models to follow prosocial values, but post-training can influence model behavior in unexpected ways. In our new paper "Predicting Alignment Generalization with Value Representations", we ask how well downstream effects of value alignment can be predicted in advance.

2173

Andy Liu @andyliu.bsky.social · 30/08/2026

many such cases (also micro vs macroeconomics to an extent)

000

Andy Liu @andyliu.bsky.social · 22/08/2026

congrats to Mythos for escaping the Sarah Chen basin

010

Reposted by Andy Liu

Ted Underwood @tedunderwood.com · 08/08/2026

initially seems good, but the more you use a tool like this, the more you lose your own ability to forecast cylones

20717113

Reposted by Andy Liu

Maarten Sap @maartensap.bsky.social · 02/02/2026

🚀 Apply to CMU LTI’s Summer 2026 “Language Technology for All” internship! 🎓 Open to pre‑doctoral students new to language tech (non‑CS backgrounds welcome). 🔬 12–14 weeks in‑person in Pittsburgh — travel + stipend paid. 💸 Deadline: Feb 20, 11:59pm ET. Apply → forms.gle/cUu8g6wb27Hs...

forms.gle

CMU LTI Summer 2026 Internship Program Application

We are looking for applicants for the Carnegie Mellon University Language Technology Institute's Summer 2026 "Language Technology for All" internship program. The main goal of this internship is to pr...

21512

Reposted by Andy Liu

Kshitish Ghate @kghate.bsky.social · 14/10/2025

🚨New paper: Reward Models (RMs) are used to align LLMs, but can they be steered toward user-specific value/style preferences? With EVALUESTEER, we find even the best RMs we tested exhibit their own value/style biases, and are unable to align with a user >25% of the time. 🧵

1127

Andy Liu @andyliu.bsky.social · 02/10/2025

Please reach out if you'd like to chat about this work! We hope ConflictScope helps researchers study how models handle value conflicts that matter to their communities. Code and data: github.com/andyjliu/con... Arxiv: www.arxiv.org/abs/2509.25369

130

Andy Liu @andyliu.bsky.social · 02/10/2025

ConflictScope can also be used to evaluate different approaches toward steering models. We find that including detailed target rankings in system prompts consistently improves model alignment with the target ranking while under conflict, but with plenty of room for improvement.

110

Andy Liu @andyliu.bsky.social · 02/10/2025

We find significant shifts between models’ expressed and revealed preferences under conflict! Models say they prefer actions that support protective values (e.g. harmlessness) when asked directly, but support personal values (e.g. helpfulness) in more realistic evaluations.

110

Andy Liu @andyliu.bsky.social · 02/10/2025

To address issues with multiple-choice evaluation, we focus on open-ended evaluation with a simulated user. Annotation studies show strong correlation between LLM and human judgments of which action a model took in a given scenario, allowing us to automate open-ended evaluations.

000

Andy Liu @andyliu.bsky.social · 02/10/2025

We introduce new metrics to measure how morally challenging a dataset is for models. We find that ConflictScope produces datasets that elicit more disagreement and stronger preferences than moral dilemma datasets, while alignment data frequently elicits indifference from models.

210

Andy Liu @andyliu.bsky.social · 02/10/2025

Given a set of values, ConflictScope generates scenarios in which an LLM-based assistant faces a conflict between a pair of values in the set. It then evaluates which value a target LLM supports more in each scenario before combining scenario-level judgments into a value ranking.

110

Andy Liu @andyliu.bsky.social · 02/10/2025

🚨New Paper: LLM developers aim to align models with values like helpfulness or harmlessness. But when these conflict, which values do models choose to support? We introduce ConflictScope, a fully-automated evaluation pipeline that reveals how models rank values under conflict. (📷 xkcd)

1164

Andy Liu @andyliu.bsky.social · 09/07/2025

Placing LLMs in simulated markets helps us quantitatively and qualitatively measure their propensity to collude, as well as how environmental changes affect this. Read below or find @veronateo.bsky.social at the ICML multi-agent systems workshop to learn more!

050

Andy Liu @andyliu.bsky.social · 06/01/2025

these are great, thanks! will check them out

000

Andy Liu @andyliu.bsky.social · 04/01/2025

started Axiomatic but didn’t get very far - Permutation City looks fun though, thanks

100

Andy Liu @andyliu.bsky.social · 03/01/2025

looking for 2025 book recs! things i've previously liked, for reference - nonfiction: the structure of scientific revolutions, cybernetic revolutionaries, seeing like a state fiction: stories of your life and others, one hundred years of solitude, project hail mary, recursion

511

Andy Liu @andyliu.bsky.social · 24/12/2024

PRISM has preference scores for different models that you can convert into pairwise labels

120

Reposted by Andy Liu

Language Technologies Institute | CMU @ltiatcmu.bsky.social · 20/11/2024

Looking for all your LTI friends on Bluesky? The LTI Starter Pack is here to help! go.bsky.app/NhTwCVb

6159

Andy Liu @andyliu.bsky.social · 07/11/2024

could I be added? thanks for curating :)

010