arXiv:2608. 15809v1 Announce Type: new Abstract: Behavioral accuracy, linear decodability, and successful activation interventions do not by themselves show that a model carries an operation-level structure from one symbolic domain to another.
arXiv:2607.27617v2 Announce Type: replace Abstract: Identical language-model answers can arise from hidden states that support different future computations, so current-answer probes do not establish...
arXiv:2608. 13626v1 Announce Type: new Abstract: A hidden state signal can be decodable or causally usable without supporting a reusable action map.
The paper introduces the concept of causal retention in interactive agents, examining whether a frozen learned state can correctly answer a mechanism‑probe map that is fixed independently of training. It shows that for finite structural causal models the optimal probe error is a Bayes decision risk, vanishing only when each learning‑interface fiber lies within a single probe‑answer fiber, and provides theoretical results such as a posterior‑coverage theorem and an exact edit decomposition. Experiments on finite causal systems, continuous simulators, TD‑MPC2, and Qwen2.5‑7B‑Instruct demonstrate that causal retention can be achieved with high accuracy, outperforming task‑performance‑based approaches.
The paper investigates PRM‑Pruned Fragment Grafting (PPFG), an inference‑time technique that extracts high‑reward prefixes from a pruned chain‑of‑thought and grafts them into a sibling decoding process. Experiments on Qwen2.5‑7B‑Instruct with Math‑Shepherd across 500 MATH problems and multiple seeds show that PPFG performs statistically indistinguishable from a parallel‑CoT baseline, with only 14% of grafts targeting genuinely struggling chains. The study extends across three language models, six benchmarks, and multiple PRM configurations, concluding that PPFG’s inertness is not due to heuristic specifics and providing an equivalence‑testing framework for mechanism nulls.
arXiv:2608. 03842v1 Announce Type: cross Abstract: When a language model fails on surface-perturbed input (typos, OCR noise, homophones), "which layer is responsible" has three natural operationalizations: where representations diverge most (sensitivity), where restoring clean activations recovers the prediction (causality), and where a small adapter can repair the damage (compensatory capacity) - and we show these three layer maps dissociate.
The paper introduces a framework for evaluating how large language model agents revise their success criteria after failures, defining five non‑compensatory conditions that must be met for a criterion revision to be considered valid. Using the CMB‑0.1 protocol, the authors test twelve cross‑domain scenarios across four system configurations, finding that no model trial satisfies all five conditions and highlighting specific failure modes such as zero‑state reconstruction and inadequate intervention sensitivity. They propose a more stringent trace‑anchored CMB‑0.4 protocol to better isolate and measure criterion revision in future studies.
The study investigates whether the reasoning steps a language model writes are causally responsible for its answers. Using a causal intervention method on the activation stream, the authors find that for Qwen3-4B, about 77% of stated steps are causally load‑bearing, while behavioral tests overestimate this by roughly 11 percentage points. The faithfulness of reasoning decreases with model size and depth of reasoning, especially for the smaller Qwen3-1.7B.
The paper introduces instruction duplication, a simple inference‑time control that repeats the procedural instruction without retraining or decoding changes. Across seven instruction‑tuned models and 16,800 scheduled generations, duplicating the instruction improves deterministic All‑8 diagnostic‑response success from 90.22% to 93.17% and reduces failures by 30.2%. In downstream Answer Engineering scenarios, duplication further boosts success rates, demonstrating its practical impact on systems that rely on the generated trajectory.
The paper presents a causal analysis of a compressed VLA policy that performs well in offline tests but fails in closed‑loop execution on a simulated pick‑and‑place task. An 8‑layer distillation of Octo‑Base retains most parameters and passes all offline metrics, yet collapses during deployment, with early stages degrading gradually and final transport failing entirely. The failure is traced to a negative, late‑heavy residual in the action trace, and standard remedies (continued training, offline data, command‑level compensation, clamping) do not restore performance; only a minimal‑pair intervention that mixes deployment‑distribution rollouts with teacher data restores parity with the teacher. whyItMatters":"The study demonstrates that offline validation metrics alone are insufficient to guarantee closed‑loop success for compressed policies, highlighting the need for targeted deployment‑time testing and interventions."
The study investigates which components of a neural network contribute to rapid generalization (grokking) and how stable that improvement remains during further training. By transferring internal attention and MLP weights along with token embeddings and readout, the authors achieve a 5.46‑percentage‑point boost in early accuracy and a 558‑step reduction in confirmation latency, while also demonstrating that freezing transferred representations largely prevents post‑grokking relapse. The work delineates clear component‑level differences between acceleration and stability, and identifies architectural limits where omitting donor embeddings leads to significant performance loss.
Instruction duplication is a simple, inference‑time control that repeats the procedural instruction in a language‑model output without retraining or decoding changes. In experiments across seven instruction‑tuned models on 300 medical multiple‑choice questions, duplicating the instruction increased the proportion of deterministic All‑8 diagnostic‑responses from 90.22 % to 93.17 % and reduced failures by 30.2 %. The technique also improved pre‑provisional TF‑IDF recall and, in downstream Answer Engineering scenarios, significantly raised success rates for specific endpoints.
arXiv:2605. 07284v2 Announce Type: replace Abstract: A late-layer change learned during post-training may work on the base model's earlier state, or it may depend on earlier computation learned with it.