iiinova — where your digital self keeps evolving

Andreas Waldis @tresiwald.bsky.social · 02/10/2026

In sum, an optimal Bayesian teacher gives LMs better, still imperfect, beliefs and a better readout than golden answers, even though it is not always right. 📄 arxiv.org/abs/2610.00679 Amazing team Polina Tsvilodub, Linlu Qiu, @tallinzen.bsky.social and Michael Franke 🙌 Questions? Reach out! (8/8)

Landing page of the project webpage, with links to the paper and code.

000 0

Andreas Waldis @tresiwald.bsky.social · 02/10/2026

BayesLM's read-out is near-optimal; OracleLM's is not (R4). In a neuro-symbolic setup, the optimal model chooses from each LM's decoded beliefs. BayesLM's accuracy barely changes (0.84 → 0.86), OracleLM's rises (0.69 → 0.78). So BayesLM's errors lie in its beliefs.

Schematic overview of the neuro-symbolic decision making pipeline, and a barplot with results below. Neuro-symbolic results of BayesLM show +0.02, neuro-symbolic results of OracleLM show +0.09.

100 0

Andreas Waldis @tresiwald.bsky.social · 02/10/2026

Bayesian fine-tuning installs better beliefs. When we patch one LM's beliefs into the other: ➔ OracleLM with BayesLM's beliefs follows Bayes more often (0.52 → 0.61) ➔ BayesLM with OracleLM's beliefs follows Bayes less often (0.77 → 0.61) (6/🧵)

Barplot showing balanced accuracy resulting from cross-model belief patching, showing that patching BayesLM into OracleLM adds +0.09, and patching OracleLM into BayesLM reduces -0.16.

100 0

Andreas Waldis @tresiwald.bsky.social · 02/10/2026

Acting on its beliefs (R3): only BayesLM passes, and only partly. When we intervene on encoded beliefs, its recommendation changes as Bayes predicts. But no single layer carries them. The edit only works across several layers at once. (5/🧵)

Schematic overview of the belief intervention and a line plot showing the balanced accuracy of BayesLM under intervention (0.77) and OracleLM (0.52) across different intervention windows.

100 0

Andreas Waldis @tresiwald.bsky.social · 02/10/2026

Acting Bayesian (R1) and holding Bayesian beliefs (R2): both LMs pass, BayesLM more so: ➔ It matches the optimal model 84% of the time vs 69%, failing mostly where the model is uncertain ➔ Its beliefs can be read from its middle layers and are closer to the model's (4/🧵)

Left: Results of evaluation of three LMs on four test sets. Middle: Simplex visualizations of LM policies over flight options of BayesLM and OracleLM. Right: Belief decoding in three LMs over the layers, where beliefs are best decoded after layer 15 in BayesLM.

100 0

Andreas Waldis @tresiwald.bsky.social · 02/10/2026

Bayesian decision-making has three layers: ➔ L3 beliefs ➔ L2 choice policy ➔ L1 action We test four requirements that map these layers to LMs. Do they: ➔ R1 act Bayesian? ➔ R2 hold Bayesian beliefs? ➔ R3 act on them? ➔ R4 compute their choice from them? (3/🧵)

Overview of the three layers of Bayesian decision-making and four formulated requirements.

100 0

Andreas Waldis @tresiwald.bsky.social · 02/10/2026

We compare two fine-tuned versions of Gemma-2 9B from Linlu Qiu et al. (2026) that recommend flights to a user with hidden preferences: ➔ BayesLM, trained on the recommendations of an optimal Bayesian model ➔ OracleLM, trained on the user's true choices (golden answers) (2/🧵)

100 0

Andreas Waldis @tresiwald.bsky.social · 02/10/2026

Is an LM that acts Bayesian also Bayesian inside? ☝️ Only as far as its beliefs allow! Fine-tuned on an optimal Bayesian model, LMs hold and use better beliefs than usual fine-tuning on golden answers. Details 👇 or bayeslm.github.io #interpretability #nlproc (1/🧵)

171 0

Andreas Waldis @tresiwald.bsky.social · 02/10/2026

In sum, an optimal Bayesian teacher gives LMs better, still imperfect, beliefs and a better readout than golden answers, even though it is not always right. 📄 arxiv.org/abs/2610.00679 Amazing team work @tsvilodub @linluqiu @tallinzen @meanwhileina 🙌 Questions? Reach out! (8/8)

000 0

Andreas Waldis @tresiwald.bsky.social · 02/10/2026

111 0

Andreas Waldis @tresiwald.bsky.social · 02/10/2026

Bar plot showing balanced accuracy from cross-model belief patching, patching BayesLM into OracleLM adds +0.09, and patching OracleLM into BayesLM reduces by 0.16.

100 0

Andreas Waldis @tresiwald.bsky.social · 02/10/2026

Acting on its beliefs (R3): only BayesLM passes, and only partly. When we intervene on encoded beliefs, its recommendation changes as Bayes predicts. But no single layer carries them: the edit only works across several layers at once. (5/🧵)

100 0

Andreas Waldis @tresiwald.bsky.social · 02/10/2026

100 0

Andreas Waldis @tresiwald.bsky.social · 02/10/2026

110 0

Andreas Waldis @tresiwald.bsky.social · 02/10/2026

We compare two fine-tuned versions of Gemma-2 9B from @linluqiu et al. that recommend flights to a user with hidden preferences: ➔ BayesLM, trained on the recommendations of an optimal Bayesian model ➔ OracleLM, trained on the user's true choices (golden answers) (2/🧵)

100 0

Andreas Waldis @tresiwald.bsky.social · 18/05/2026

The same encode-vs-express gap also shows up for factual knowledge🙌 bsky.app/profile/zori...

020 0

Andreas Waldis @tresiwald.bsky.social · 18/05/2026

ICL mostly changes how LMs produce output tokens from input information that stays largely stable regardless of the instruction. So unwanted behavior, a failure in output token distributions, doesn't necessarily mean the model isn't encoding what we'd expect. (1/2)

120 0

Andreas Waldis @tresiwald.bsky.social · 18/05/2026

More on this production-centered mechanism across models and + implications for evaluation, interpretation, and pre-training: 🔗 instruction-probing.github.io 📄 arxiv.org/abs/2605.11206 Team effort with @lchoshen.bsky.social @yufanghou.bsky.social @yperlitz.bsky.social🙌 Questions? Reach out! (3/3)

Project picture with icon, title, and contributors

010 0

Andreas Waldis @tresiwald.bsky.social · 18/05/2026

In short, instructions act less on what models process, and more on what they emit. Behavior changes from prompting, including prompt instability and in-context learning, therefore seem to arise mainly at the production stage, with little adaptation during task-sample processing. (2/🧵)

110 0

Andreas Waldis @tresiwald.bsky.social · 18/05/2026

Do instructions affect how LMs process and produce language? ☝️Not the way you think! 😲LMs barely change task information when processing a task sample. Instead, instructions shape how this information is accessed and expressed when producing output tokens. #interpretability #nlproc (1/🧵)

1141 0

Reposted by Andreas Waldis

Vagrant Gautam @dippedrusk.com · 27/03/2026

I already presented some work on reference (names, pronouns, coreference resolution, pronoun fidelity, etc.) as a rich site to evaluate biases and commonsense reasoning, and our work on disentangling model behaviour and internals through aligned probing (led by @tresiwald.bsky.social).

151 0

Andreas Waldis @tresiwald.bsky.social · 25/03/2026

Excited to present this work together with @dippedrusk.com at #EACL. Join us in the poster session 1 (11:30-13:00) 🔥

041 0

Andreas Waldis @tresiwald.bsky.social · 03/03/2026

Thanks a lot to everyone for the support, guidance, mentoring, collaboration, and great moments over the past years! 🙏 Without you, this journey wouldn't have been such a pleasure — and now excited to see what the future brings! 🚀

151 0

Andreas Waldis @tresiwald.bsky.social · 27/01/2026

Bonus: 4 case studies in the paper 💠 DPO detoxification impacts upper layers but loses general info 💠 Toxicity varies across prompt formulations, but internals stay stable 💠 Insights hold under quantization 💠 Encoding patterns emerge early in pre-training (6/🧵)

100 0

Andreas Waldis @tresiwald.bsky.social · 27/01/2026

Connecting behavioral and internal results shows that more info about input toxicity correlates with less toxic outputs. With layer-wise intervention, we show this effect is causal—layer skipping increases toxicity up to +0.16. (5/🧵)

analysis of the correlation between toxic behavior and information about toxicity and causal evidence from skipping model layers

120 0

Andreas Waldis @tresiwald.bsky.social · 27/01/2026

Instruction-tuning makes LMs less toxic: 📈 More info about input toxicity 📉 Less info about output toxicity Strongest for contextual dimension like Threat, thus instruction-tuning seems to impact semantics, not just keywords. (4/🧵)

comparison of how pre-trained and instruction-tuned LMs encode toxicity internally

110 0

Andreas Waldis @tresiwald.bsky.social · 27/01/2026

We show where LMs encode toxicity: 💠Lower layers encode most info 💠Output toxicity detectable in input tokens 💠Input toxicity propagates to output 💠Context-dependent dimensions (e.g., Threat) peak in higher layers than word-sensitive ones (e.g., Sexually Explicit) (3/🧵)

an overview of the four probing scenarios we use to measure toxicity information within LMs

110 0

Andreas Waldis @tresiwald.bsky.social · 27/01/2026

Across 6 LMs: 📈 Outputs +0.27 more toxic than human continuations 📈 Input-output toxicity correlation is stronger for both toxic (+0.30) and non-toxic (+0.32) prompts LMs replicate and amplify what they're fed. (2/🧵)

table of evaluation results of the toxic behavior across six toxicity dimensions

110 0

Andreas Waldis @tresiwald.bsky.social · 27/01/2026

LMs that "know more" about toxicity are less toxic! Our #TACL 📄 connects behavior and internals: 💠 LMs amplify toxicity beyond humans 💠 Information about toxicity peaks in lower layers 💠 Bypassing these layers increases toxicity More details👇 #NLProc #interpretability (1/🧵)

simplified overview of our aligned probing setup, where we join the behavioral and internal evaluation of LMs' toxicity

1167 0

Reposted by Andreas Waldis

INTERPLAY Workshop@COLM '25 @interplay-workshop.bsky.social · 09/10/2025

✨ The schedule for our INTERPLAY workshop at COLM is live! ✨ 🗓️ October 10th, Room 518C 🔹 Invited talks from @sarah-nlp.bsky.social John Hewitt @amuuueller.bsky.social @kmahowald.bsky.social 🔹 Paper presentations and posters 🔹 Closing roundtable discussion. Join us in Montréal! @colmweb.org

Schedule for the INTERPLAY workshop at COLM on October 10th, Room 518C.

09:00 am: Opening
09:10 am: Invited Talks by Sarah Wiegreffe and John Hewitt
10:20 am: Paper Presentations

Lunch Break

01:00 pm: Invited Talks by Aaron Mueller and Kyle Mowhald
02:10 pm: Poster Session
03:20 pm: Roundtable Discussion
04:50 pm: Closing

044 0

Reposted by Andreas Waldis

INTERPLAY Workshop@COLM '25 @interplay-workshop.bsky.social · 08/07/2025

Missed a spot? If you have a pre-reviewed paper from ARR or COLM that focuses on the INTERPLAY between LM internals and behavior, there is a shortcut to presenting at our @colmweb.org workshop! ✨ Join us in Montréal! 🇨🇦 CfP: shorturl.at/sBomu OpenReview: shorturl.at/WwWhg #nlproc #interpretability

Call for Pre-Reviewed Papers, Interplay Workshop at COLM: July 10th - submissions due. July 24th - acceptance notification. October 10th - workshop day.

152 0

Reposted by Andreas Waldis

INTERPLAY Workshop@COLM '25 @interplay-workshop.bsky.social · 24/06/2025

Delighted that ✨Mor Geva (@megamor2.bsky.social) and ✨Anna Ivanova (@neuranna.bsky.social) will complete our speaker line-up and talk about the INTERPLAY of model internals and behavior. Be there and submit by June 30th 📄 shorturl.at/sBomu See you in 🇨🇦 @colmweb.org #nlproc #interpretability

Mor Geva and Anna Ivanova will talk at the INTERPLAY workshop.

053 0