Recent

Discovery surface for what's new in the corpus. Time-windowed view derived from created_at on every table — papers, restate edges, cross-corpus bridges, communities, and god-node movers. Pick a window:

New papers (36)

Santiago Aranguri

Rates of rare harmful behavior in Qwen 3 4B can be estimated with up to 30× fewer rollouts than naive sampling by borrowing statistical strength from a less-safe sibling model. Logit Path Extrapolation interpolates between Qwen 3 4B and Qwen 3 4B Abl…

0 communities·9 claims·8 findings
Santiago Aranguri

A training-free logit monitor, which token-forces a prespecified eval-aware sentence and multiplies per-token probabilities at every CoT sentence boundary, separates evaluation from deployment prompts better than an LLM-judge verbalization monitor wh…

0 communities·11 claims·18 findings

A sparse autoencoder trained on layer-19 residual-stream activations of Llama-3-8B-Instruct, using LMSYS-Chat-1M, yields features that can be labeled automatically, ranked by causal relevance, and used to steer chat behavior, and training-data choice…

0 communities·11 claims·11 findings
Thomas McGrath

Sparse autoencoders trained on an intermediate layer of Llama 3.3 70B, using LMSys-Chat-1M data, yield a browsable, steerable feature space, and the clearest results are about where it breaks: steering works but silently degrades facts, and LLM-judge…

0 communities·8 claims·8 findings
Dron Hazra

Reasoning models route around interventions that work on standard LLMs: in DeepSeek R1 (671B parameters), steering fails at the first token of the chain of thought, and pushing a steered feature harder first removes the intended behavior before the o…

0 communities·6 claims·9 findings

Cross-Layer Transcoders (CLTs) trained on GPT-2 Small recover interpretable greater-than machinery from a single prompt, but the recovered circuit is messier and more lookup-like than the MLP-neuron circuit mapped by Hanna et al. (2023). The CLT used…

0 communities·10 claims·16 findings
Thomas McGrath

Sampling from logits extrapolated past the post-trained model, logits_after + α(logits_after − logits_before), surfaces rare undesired behaviors 10 to 300 times more often than standard sampling, with samples that stay coherent. Model diff amplificat…

0 communities·11 claims·11 findings
Gorton, Liv · Lewis, Owen (2025)

Superposition causally contributes to adversarial vulnerability, and the link runs in both directions: in a toy ReLU autoencoder (n=100 features, m=20 hidden units), raising feature sparsity raises features-per-dimension and adversarial vulnerability…

0 communities·10 claims·12 findings
Yoav Gur-Arieh · Mor Geva · Atticus Geiger (2025)

Language models retrieve bound entities through a mixture of three mechanisms, not the positional mechanism alone, and the positional signal collapses in the middle of long entity lists. With n=20 entity groups on gemma-2-2b-it, positional patching c…

0 communities·9 claims·28 findings
Lucius Bushnaq · Dan Braun · Lee Sharkey (2025)

Rank-one subcomponents trained against stochastic ablation masks recover ground-truth mechanisms in toy networks that Attribution-based Parameter Decomposition (APD) could not decompose, while removing APD's top-k hyperparameter, gradient attribution…

0 communities·10 claims·15 findings
Merullo, Jack · Vatsavaya, Srihita · Bushnaq, Lucius (2025)
0 communities·10 claims·18 findings
Aaditya Vikram Prasad · Connor Watts · Jack Merullo (2026)

A frozen Gemma-3-12B-IT's own activations can serve as a dense reward for open-ended factuality training, replacing a $344,064 Gemini 2.5 Pro web-search judge with a roughly $3,818 probe pipeline. The method, RLFR (Reinforcement Learning from Feature…

0 communities·15 claims·29 findings
Jing Huang · Daniel Wurgaft · Rachit Bansal (2026)

Larger models learn tasks smaller models never learn, and the cause is data-induced competition for neurons, not expressivity: small models spend capacity on frequent or simple tasks, and their updates overwrite rare-task features before the next rar…

0 communities·13 claims·22 findings
Eric Bigelow · Raphaël Sarfati · Daniel Wurgaft (2026)

Llama-3.1-8B-Instruct's in-context belief updates while reading a story trace smooth trajectories on a low-dimensional manifold. That manifold is shared between behavior and residual activations, and its geometry predicts how steering spills across c…

0 communities·10 claims·20 findings
Usha Bhalla · Thomas Fel · Can Rager (2026)

Sparse autoencoders trained on Llama3.1-8B (layer 19, 500M Pile tokens, 25 SAEs across Standard, JumpReLU, TopK, BatchTopK and Matryoshka architectures at VE > 0.85) do not capture continuous concept manifolds as compact subspaces; they spread them a…

0 communities·18 claims·19 findings
Thomas Fel · Matthew Kowal · Mozes Jacobs (2026)

Lifting the featurizer's atom from a direction to a block of directions recovers concept manifolds that SAEs shatter, and block sparsity is the MAP-matched prior for a representation that is a sparse Minkowski sum of low-dimensional manifolds. The pa…

0 communities·19 claims·20 findings
Leon Bergen · Usha Bhalla · Sidharth Baskaran (2026)

Interpretability can audit a preference dataset before optimization and predict what DPO will teach a model: on Dolci, clustering SAE features and testing chosen-minus-rejected activation differences predicted post-DPO feature shifts with R²=0.9 (fea…

0 communities·17 claims·23 findings
Leon Bergen · Usha Bhalla · Andrew Lee (2026)

Difference-of-means (DoM) vectors built from 5,285 templated synthetic SWE conversations detect reward hacking in long agentic rollouts of Kimi K3, GLM 5.2, and Qwen 3.8 Max, matching a GPT-5.6 Sol LLM monitor at near-zero inference cost. The authors…

0 communities·11 claims·23 findings
Eric Bigelow · Amir Zur · Satchel Grant (2026)

Resampling noise in reasoning-chain uncertainty dynamics is almost exactly multinomial, so most of the cost of Forking Paths Analysis re-measures structure that neighboring tokens already contain. On tinyMMLU (n=100) with Llama-3-8B-Instruct and Deep…

0 communities·7 claims·26 findings
Nicholas L. Turner · Jeffrey Wu · Joshua Batson

Interference weights are measurable inside a trained transformer, and the model keeps its most effective virtual weights helpful while confining harmful ones to a low-effectiveness range. In a one-layer, 2.9M-parameter model (d_m=256, 4,096-token voc…

0 communities·12 claims·18 findings
Trenton Bricken · Rowan Wang · Sam Bowman (2025)

LLM auditing agents can recover a planted hidden objective at rates that are modest individually but strong in aggregate: a Claude Sonnet 4 tool-using investigator solves the Marks et al. reward-model-sycophancy auditing game 13% of the time under re…

0 communities·11 claims·24 findings
Michaud, Eric J. · Gorton, Liv · McGrath, Tom (2025)

Sparse autoencoder scaling laws can break down when activations contain feature manifolds: if per-feature loss falls slowly with latent count, an SAE spends its latents tiling common manifolds instead of discovering rarer features. Michaud, Gorton an…

0 communities·9 claims·13 findings

Attribution graphs built on transcoders froze attention patterns, so they could not say why a head attended where it did, and for many prompts that missing explanation is the whole computation. Tracing Attention fixes this with QK attributions, which…

0 communities·11 claims·19 findings
Chris Olah

Transcoders can learn a different computation from the layer they replace, even when the layer is trivial and the transcoder is monosemantic with low on-distribution error. In an absolute-value toy model, y = ReLU(x) + ReLU(-x), a transcoder trained …

0 communities·9 claims·7 findings
Kevin Der · Harish Kamath · Thompson, Ben (2026)

Training a sparse autoencoder on the mean residual-stream activation of a conversation turn, rather than on individual tokens, yields features that describe a turn's high-level properties more completely than per-token features do, and makes attribut…

0 communities·11 claims·35 findings
Thomas Jiralerspong · Trenton Bricken (2026)

Crosscoders can diff models with different architectures, and a modified crosscoder, the Dedicated Feature Crosscoder (DFC), surfaces behaviorally meaningful model-exclusive features without supervision. The DFC hard-partitions the 131,072-feature di…

0 communities·14 claims·31 findings
Siddharth Mishra-Sharma · Trenton Bricken · Jack Lindsey

Crosscoder model diffing produces model-exclusive features that are systematically denser and more polysemantic than shared features, and this is an artifact of feature-budget competition that a small change to the loss largely removes. In smaller Cl…

0 communities·8 claims·13 findings
Adam Karvonen · Can Rager · Johnny Lin (2025)

Gains on the sparsity-fidelity frontier do not predict how well a sparse autoencoder performs on downstream tasks, and SAEBench demonstrates this by scoring over 200 SAEs on eight metrics. Matryoshka BatchTopK SAEs rank below TopK and BatchTopK on lo…

0 communities·12 claims·26 findings
Adam Karvonen · Euan Ong · Subhash Kantamneni (2026)

Activation-reading interpretability tools give no uplift for predicting counterfactual model behavior on naturally occurring behaviors, even though the same tool families help in narrow fine-tuned-quirk auditing games. CHIVE (Counterfactual Hypothesi…

0 communities·10 claims·25 findings
Authors

A single attention head's behavior on the full data distribution rarely matches what a narrow task suggests, and HeadVis, an interactive tool for browsing heads in Claude Haiku 3.5 and Gemma 3 1B, makes that gap visible. HeadVis combines a metric sca…

0 communities·15 claims·21 findings
Authors

A sparse, privileged subset of residual-stream directions in Claude models meets five functional criteria of a global workspace, and it is found by asking which representations the model is poised to verbalize. The Jacobian lens (J-lens) averages the…

0 communities·27 claims·61 findings
Shamil Chandaria · Arvo Muñoz Morán · Fernando Rosas (2026)

Five levels of functional description — behavioural, computational, intrinsic causal-structural, organismic, and organism-environment — collapse the theoretical cacophony of consciousness science into a taxonomy where the deepest disagreement about A…

0 communities·19 claims·12 findings
Susumu Inui (2026)
0 communities·16 claims·0 findings
Samson Abramsky · Wolfgang Banzhaf · Leo S. D. Caves (2026)
0 communities·15 claims·4 findings
Jeffrey Lai · Anthony Bao · J. Quinn (2026)
0 communities·14 claims·17 findings

God-node movers (20)

Entities that gained the most new edges. Often signals "this thinker / framework / community just got reinforced by fresh material."

New cross-paper restate edges (25)

Claims/findings/hypotheses in different papers that paraphrase each other (cosine ≥0.90). New restates often signal "the corpus just got two papers making the same claim — that claim is becoming consensus" or "fresh contradiction detected."

New cross-corpus bridges (0)

External markdown (aboutblank KB, Alexander notes, Zen notes, research notes) newly linked to corpus entities via Nomic cosine. High-cosine bridges are essay-candidate seeds.

No new bridges.

New communities (0)

Clusters formed by the weekly Leiden detector. New communities often signal "a fresh theme has enough material to form a cluster."

No new communities.