Skip to main content
Search

# Linear Probing

10 related articles

Subscribe to updates
  1. bioRxiv

    A frozen EEG foundation model needs no fine-tuning: one first-layer attention head's connectivity matrix reaches 42.13 balanced accuracy in cross-subject emotion recognition

    The EEG foundation model REVE no longer needs full fine-tuning or all-layer pooled features to recognize emotion. Such models had not consistently beaten task-specific architectures on cross-subject tasks, and their large representations are costly to compute and store. This approach aggregates the first-layer attention weights of a single head over each trial into a weighted channel-to-channel matrix, then uses a label-free symmetry score to pick the most informative head; on SEED-IV, classical classifiers reach 42.13 balanced accuracy with about 176 times fewer calculations than all-layer attention pooling.

  2. arXiv

    Video language models lose count of repeated actions: answers stick to a few favorite numbers while the count stays in the last layer

    RepProbe turns repetition count into a controlled variable and finds that generalist video language models collapse at high counts even though the count remains linearly decodable from the same final-layer state, locating the failure in the output pathway rather than the representation.

  3. arXiv

    Wearable pulse-wave foundation models meet a common exam: 35.3% of 68 task-regime combinations fail to beat a random baseline

    An open benchmark evaluates five public PPG foundation models across nine datasets, 26 tasks and four downstream adaptation regimes under one protocol, finding that 35.3% of 68 task-regime combinations show no meaningful gain over a random or constant baseline and 83.8% show no PPG-specific model beating both classical pulse-rate-variability features and a generic time-series foundation model.

  4. arXiv

    Brain MRI foundation models lose accuracy on a new dataset: not because they remember the source, but because each dataset encodes the sequence along different directions

    Across 17 public brain MRI datasets and eight frozen representations, the sequence is recognizable inside every dataset yet collapses on an unseen one, because each dataset encodes the sequence difference along different directions rather than because the features remember where a scan came from.

  5. arXiv

    Post-training slices a language model's belief state into a causal quotient: reward-null information is mostly rerouted or rescaled, and only sustained weight decay erases it

    In hidden-Markov worlds where pretraining recovers Bayesian belief states, the authors define an exact reward-null kernel from a reward that reads only a coarse function of the hidden state, and separately measure whether reward-null information stays linearly decodable, whether decisions causally depend on it, and how much activation variance it occupies; post-training mostly reroutes or rescales this information while leaving it decodable, a KL anchor keeps decisions using it, unanchored objectives let decisions stop using it, and only prolonged weight decay erases distinctions that neither reward nor next-token prediction can see, with open language models showing the same dissociation in in-context belief geometry.

  6. arXiv

    FairRSFM evaluates three remote sensing foundation models by biome group and finds aggregate metrics mask ecological performance gaps

    The authors introduce FairRSFM, which maps georeferenced samples from four remote sensing downstream datasets (m-EuroSAT, m-BigEarthNet, m-SA-Crop-Type, and MMEarth20K) to 14 terrestrial biome classes consolidated into six macro-groups, and evaluates Prithvi-EO-2.0, SatMAE, and DOFA under a unified frozen-backbone protocol across three random seeds, showing that aggregate metrics consistently mask biome-dependent disparities, for example Prithvi-EO-2.0 reaching high overall macro-F1 on m-EuroSAT while its worst-group score is markedly lower, and compares BOLP, DBR, and GroupDRO as mitigation baselines whose effectiveness is model- and task-dependent.

  7. arXiv

    EscapeGuard cuts tool-selection hallucination by 9.0 pp across six benchmarks and six models while suppressing cross-configuration hallucination escape

    The work identifies and formally defines hallucination escape, in which existing tool-hallucination mitigations reduce hallucination on the tool configuration they are tuned on but increase it on others; it finds that models hold intrinsic tool-use tendencies whose conflict with the runtime configuration sharply raises hallucination and that existing methods reinforce these tendencies; it then proposes EscapeGuard, a training-free inference-time method combining conflict-aware gating with configuration-derived attention enhancement, which reduces tool-selection hallucination by 9.0 pp, lowers the cross-configuration mean by 23.7 pp, and achieves 89.1% net improvement in paired-query evaluation across six benchmarks and multiple models.

  8. arXiv

    No Dedicated Perceptual Encoder Needed: Ogata and Nakashima Find Multimodal Transformers Grow a "Virtual Encoder" in Early-to-Middle Layers

    Ogata and Nakashima ask where encoding happens when multimodal language models skip dedicated perceptual encoders and expose a shared transformer to lightly projected patches, audio frames, or discrete visual tokens, and through linear probing, similarities to perceptual encoders, and causal analyses they find the transformer internalizes the missing computation, constructing task-usable perceptual representations within its own early-to-middle layers, a structure they call a Virtual Encoder.

  9. arXiv

    Tri-PvP's 8,000 tri-modal conflict samples show omni-modal LLMs favor images, and favor perceptual vision but propositional audio

    The authors built Tri-PvP, an 8,000-sample tri-modal conflict benchmark spanning animal, emotion, environment, and music domains in which image and audio each appear in perceptual form (real photographs or recordings) or propositional form (text-card images or TTS speech) while text is always propositional, evaluated five omni-modal LLMs across four evidence-type conditions, and found that image bias dominates in 18 of 20 model-by-condition bars and that a systematic asymmetry exists: models favor perceptual evidence in vision but propositional evidence in audio, with this bias linearly decodable from early hidden layers and only partially mitigated by contrastive decoding, which introduces a residual text bias.

  10. medRxiv

    Task-Specific Quality Gating for Retinal OCT B-Scans: Learned Representations Over Scalar Metrics in Choroid Segmentation

    Using choroid segmentation as a prototype task on 6,076 OCT B-scans from 80 subjects, this study systematically compared scalar no-reference image quality metrics (BRISQUE, NIQE, PIQE, SNR, PSNR), general-purpose ImageNet-pretrained representations, and retinal foundation models as task-specific quality gates, finding that scalar metrics correlate weakly with segmentation Dice (|r| < 0.20), that general-purpose pretrained representations reach linear-probe ROC-AUC up to about 0.77, that the OCT-specific foundation model RETFound reaches about 0.81, and that only the retinal foundation model embeddings form quality-aligned unsupervised K-Means clusters exceeding a patient-level permutation null.

You’re all caught up