finding

active

finding:mds-injections-outperform-p2-in-open-ended-generation-in-11-of-14-llms-with-phi-gains-of-3-61-to-16-44

MDS injections outperform P2 in open-ended generation in 11 of 14 LLMs with Phi gains of 3.61% to 16.44%

Primary quantitative result overturning prior reports that prompting outperforms representation engineering

Source paper

extracted_from

Psychological Steering of Large Language Models

(2026) · Leonardo Blas · Robin Jia · Emilio Ferrara

Neighborhood — ranked by edge-count

Papers (1)

paper

Psychological Steering of Large Language Models
introduces

Thinkers (1)

thinker

Amin Banayeeanzade
contradicts
Lead author of prior work reporting P2 outperforms MD injections; the paper overturns this result

Claims (2)

claim

Uncalibrated sweep units and restricted coefficient ranges are the primary cause of prior reports showing P2 outperforming MD injections
associated_withsupports
Mechanistic explanation for discrepancy with Banayeeanzade et al.; addressed by centroid unit and unbounded sweep contributions
RepE is a new frontier in open-ended psychological steering of LLMs, outperforming prompting when properly calibrated
associated_withsupports
Central interpretive claim overturning prior reports; supported by 11-of-14 LLM wins for MDS over P2

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

PM hybrid outperforms both P2 and MDS in 13 of 14 LLMs with Phi gains over P2 from 5.56% to 21.92% and over MDS from 3.30% to 26.67%finding0.879
Key finding showing that combining prompting and injection is the strongest approach
MDS achieves global win proportion of 89.5% on SJTs across 14 LLMs and four injection stridesfinding0.806
MDS dominates in open-ended generation by global win proportion metric (Table 2)
MDS achieves global win proportion of 47.3% on MPI-120 inventory across 14 LLMsfinding0.789
MDS is also the top method on the inventory task but with much smaller margin than on SJTs (Table 2)
MDS injections align with the Linear Representation Hypothesis: target trait varies near-linearly with alpha in open-ended generationclaim0.774
Theoretical alignment claim backed by OLS R2 analysis showing 96.15% of trends have R2>=0.75
MDS injections show no salient patterns in MPI-120 inventory responses beyond occasional co-occurring peaksfinding0.773
Contrasts with SJT results; leads authors to narrow analyses to SJT responses
gemma-3-1b-it yields only one valid MDS injection score (phi_1,A,up = 4.8) and is excluded from main analysesfinding0.764
Identified exception to overall MDS effectiveness; reason remains unexplained as a limitation
MDS injection steering efficiency peaks at mid-layers across LLMs, injection strides, and OCEAN traitsfinding0.752
Consistent empirical pattern supporting the connection between mid-layer representations and emotion/behavioral content
Why do MDS injections outperform other methods on the inventory (multiple-choice) task?question0.752
Identified as an unexplained result and open question in limitations section