new

Get trending papers in your email inbox!

Subscribe

Daily Papers

byAK and the research community

Nov 18

Decoding specialised feature neurons in LLMs with the final projection layer

Large Language Models (LLMs) typically have billions of parameters and are thus often difficult to interpret in their operation. Such black-box models can pose a significant risk to safety when trusted to make important decisions. The lack of interpretability of LLMs is more related to their sheer size, rather than the complexity of their individual components. The TARS method for knowledge removal (Davies et al 2024) provides strong evidence for the hypothesis that that linear layer weights which act directly on the residual stream may have high correlation with different concepts encoded in the residual stream. Building upon this, we attempt to decode neuron weights directly into token probabilities through the final projection layer of the model (the LM-head). Firstly, we show that with Llama 3.1 8B we can utilise the LM-head to decode specialised feature neurons that respond strongly to certain concepts, with examples such as "dog" and "California". This is then confirmed by demonstrating that these neurons can be clamped to affect the probability of the concept in the output. This extends to the fine-tuned assistant Llama 3.1 8B instruct model, where we find that over 75% of neurons in the up-projection layers have the same top associated token compared to the pretrained model. Finally, we demonstrate that clamping the "dog" neuron leads the instruct model to always discuss dogs when asked about its favourite animal. Through our method, it is possible to map the entirety of Llama 3.1 8B's up-projection neurons in less than 15 minutes with no parallelization.

  • 1 authors
·
Jan 5, 2025

Hydra: Unifying Document Retrieval and Generation in a Single Vision-Language Model

Visual document understanding typically requires separate retrieval and generation models, doubling memory and system complexity. We present Hydra, a dual-head approach that provides both ColBERT-style late-interaction retrieval and autoregressive generation from a single vision-language model (VLM). A single LoRA adapter, trained only for retrieval, is toggled at inference: enabling it produces multi-vector embeddings; disabling it recovers the base model's generation quality -- byte-identical outputs in 100% of 10,500 greedy and stochastic samples, with max delta-ANLS = 0.0044 across 15,301 samples on four VQA benchmarks (three informative; ChartQA is near-zero for both models under greedy decoding) when compared against an independent base-model pipeline. We identify three engineering requirements (attention-mode restoration, lm_head preservation, KV-cache-aware decoding) whose omission silently breaks generation despite correct weight recovery. On ViDoRe V1, Hydra (4B) is within 1 percentage point of a controlled single-head baseline in a single training run, with higher aggregate scores on V2 and V3 that are concentrated on a subset of tasks; multi-seed experiments are needed to confirm these trends. The single-model design reduces peak GPU memory by 41%, though adapter switching introduces throughput overhead under concurrent serving loads. An ablation shows that GritLM-style joint training provides no benefit within the LoRA-based (r=16) training regime. A proof-of-concept extension to Qwen2.5-Omni-3B demonstrates that the mechanism generalizes to audio retrieval and video embedding, with speech generation.

  • 1 authors
·
Mar 30

Bone: Block Affine Transformation as Parameter Efficient Fine-tuning Methods for Large Language Models

Low-Rank Adaptation (LoRA) has achieved remarkable training results by freezing the original weights and training only low-rank matrices, establishing itself as the predominant fine-tuning method for LLMs. In pursuit of performance closer to full-parameter training, a series of LoRA variants have emerged, such as LoRA+, PISSA, Olora, and LoRA-GA. However, these improvements complicate the initial setup of model training and increase initialization time. More importantly, they overlook the internal interactions of the original weight information. To address these issues, we introduce a novel theory, ``Weight Guide'' aimed at continuously guiding trainable matrices through the original weights during training to enhance the utilization of weight information. Based on this theory, we designed a new PEFT technique called Bone (Block Affine), which not only enhances the utilization of original weight information but also emphasizes the internal connections between weights, leading to faster convergence and better data fitting. Experimental comparisons across two different LLM architectures (LLaMA2, RWKV6) and various parameter scales demonstrate that the Bone structure can achieve rapid convergence and superior data fitting without the need for complex initialization. For example, when fine-tuning LLaMA2-7B on the MetaMathQA dataset and validating on GSM8k and math benchmarks, Bone achieved fine-tuning scores of 49.36 and 8.8, respectively, outperforming PISSA by 5.84\% and 1.96\%.

  • 1 authors
·
Sep 19, 2024 1

Can an LLM Actually Reason and Program about Neural Network's Weights or is it just memorization on steroids? Exploring "Weight-Space Programming"

Large language models (LLMs) have recently been reported to read and write the raw weights of neural networks directly, a capability termed weight-space programming. We ask whether the reading side of this capability reflects reasoning about weight flows or a combination of local arithmetic at the network's edges and memorized priors over public datasets. We test Claude Opus 5.5 on small convolutional networks of about 125,000 parameters trained on five image datasets, providing every weight as plain text with no tools, no forward passes, and no dataset or class names. The model performs accurate local calculations: batch-normalization statistics lead it to infer about three optimizer steps on networks stopped after three updates, and output-layer sign patterns recover the vehicle/animal split of CIFAR-10. The sessions report reading the first and last layers closely and sampling the middle, and several claims built on partial tracing fail verification, although we did not run a direct forward-pass test. Semantic judgments lean on familiar datasets: the model names CIFAR-10 for CIFAR-10 networks but also for all three fast-food networks, relabels KMNIST as MNIST, reads a 99.57%-accurate SAT-6 network as CIFAR-10 animal photos, and calls "frog" the output favored by a feature that fires most on trees. Permuting CIFAR-10's labels preserves the output split at the new indices, but every class pair the model interprets is then wrong, indicating reliance on remembered label order. These are exploratory results from mostly single runs. We conclude that zero-shot weight reading by this model is better described as local arithmetic plus dataset priors than as weight-flow tracing, and we provide a protocol for auditing future weight-space claims.

  • 1 authors
·
Nov 4

Does RoPE Prevent or Degrade Retrieval Heads? A Mechanistic Analysis Across Model Families

Retrieval heads, attention heads that copy information from earlier context to the current position, have been proposed as the mechanistic substrate for long-context recall. Rotary position embeddings (RoPE) rotate queries and keys by frequencies decaying with a base hyperparameter theta, and a natural hypothesis is that this rotation either prevents retrieval heads from forming or degrades their function. We test both across four open-weight 7-8B models spanning multi-head and grouped-query attention and a 100x range of theta, using paired-seed needle-in-a-haystack tests, layer-clustered permutation, and causal head-masking. (i) Retrieval heads are causally necessary: masking the 87 detected heads in OLMo-2 collapses recall from 1.00 to 0.00, while masking matched random heads has no effect; this replicates in Qwen. (ii) Higher theta does not reduce retrieval-head count (LLaMA-3.1 at theta=500K has 47 heads vs LLaMA-2 at theta=10K with 42), refuting the prevention hypothesis. (iii) The norm-utility relation is family-specific and significant in opposite directions (Qwen d=-0.49, OLMo d=+0.50, both significant; LLaMA null); since OLMo and LLaMA-3.1 share theta=500K yet differ, the effect is not theta-driven. (iv) Building on Chiang and Yogatama (2025), a controlled patch shows that zeroing the lowest-frequency RoPE dimensions of retrieval heads degrades recall dose-dependently (1.00 to 0.18 when 32 of 128 dimensions are zeroed, vs 0.98 for random dimensions); the effect is head-specific and task-specific. The causal variable is RoPE frequency, not norm-utility. The direction holds in all five models patched (OLMo-2, Qwen2.5-7B/14B, Gemma-2, Mistral) across four lineages and two scales. We do not claim cross-model magnitude. Code and a paired-seed harness are released.

  • 1 authors
·
Jun 18

Understanding Linear Probing then Fine-tuning Language Models from NTK Perspective

The two-stage fine-tuning (FT) method, linear probing (LP) then fine-tuning (LP-FT), outperforms linear probing and FT alone. This holds true for both in-distribution (ID) and out-of-distribution (OOD) data. One key reason for its success is the preservation of pre-trained features, achieved by obtaining a near-optimal linear head during LP. However, despite the widespread use of large language models, there has been limited exploration of more complex architectures such as Transformers. In this paper, we analyze the training dynamics of LP-FT for classification tasks on the basis of the neural tangent kernel (NTK) theory. Our analysis decomposes the NTK matrix into two components. This decomposition highlights the importance of the linear head norm alongside the prediction accuracy at the start of the FT stage. We also observe a significant increase in the linear head norm during LP, which stems from training with the cross-entropy (CE) loss. This increase in the linear head norm effectively reduces changes in learned features. Furthermore, we find that this increased norm can adversely affect model calibration, which can be corrected using temperature scaling. Additionally, we extend our analysis with the NTK to the low-rank adaptation (LoRA) method and validate its effectiveness. Our experiments using a Transformer-based model on multiple natural language processing datasets confirm our theoretical analysis. Our study demonstrates the effectiveness of LP-FT for fine-tuning language models. Code is available at https://github.com/tom4649/lp-ft_ntk.

  • 2 authors
·
Oct 21, 2024

The strength of clinical evidence is recoverable from language model representations but not from their stated grades

Large language models (LLMs) increasingly summarize clinical evidence, where a claim's weight depends on how strongly it is supported. Yet these models convey confidence poorly, and properties they never state, such as truth, are often readable from their activations. Whether a clinical model registers evidence strength, distinct from truth, and states it when asked is untested, and any such signal could be lexical. We compiled 45,134 clinical claims from six public sources, harmonized 20,611 into a four-level evidence grade under three independent frameworks, and tested 22 local, open-weight LLMs from several developers (0.6-70 billion parameters; general, medical, and reasoning), with lexical, truth, and cross-framework controls. A linear estimator recovered the grade in every model (median AUROC 71.8), yet decodability did not rise with scale and was weakest in reasoning models. The grade the models stated fell to chance, 25-27 percentage points below the estimator. The recoverable signal was largely lexical and did not transfer across topics or frameworks, yet it was distinct from factual truth and still flagged weakly supported claims (AUROC 69.2). Clinical LLMs thus carry an ordered evidence-strength signal they do not express, so their stated grades fail to convey a claim's support even when it is recoverable from their representations and text.

  • 1 authors
·
Jun 26