BACH-V: Bridging Abstract and Concrete Human-Values in Large Language Models

AI summary

Overview

• Research area: Natural Language Processing / LLM interpretability and value alignment (mechanistic probing and activation steering).
• Technical level: Advanced.
• Scope: This paper introduces an "abstraction-grounding" framework that uses linear probes and activation steering to test whether six open-source LLMs internally represent ten human values and whether those representations transfer from abstract concept descriptions to concrete events and decisions.

What This Paper Is About

The central question is whether large language models genuinely understand abstract concepts such as justice, curiosity, or individualism, or whether they merely manipulate these terms as statistical patterns. The authors use human values as a testbed because values are semantically rich and central to alignment, and they separate "understanding" into three distinct capacities — interpreting abstract concepts (Abstract-Abstract, A-A), grounding abstractions in concrete events (Abstract-Concrete, A-C), and applying abstract principles to regulate decisions (Concrete-Concrete, C-C). The goal is to establish whether value representations exist in models, whether they are causally effective, and whether they transfer across these levels.

Key Contributions

1. A concept probing toolkit that reads internal LLM representations and shows that distinct value concepts produce distinguishable neural signatures, establishing that value representations exist.
2. A behavior steering pipeline that modifies internal representations along the probe direction and produces predictable, controllable behavioral shifts, establishing causality.
3. A systematic evaluation system spanning the three interaction regimes (A-A, A-C, C-C), crossed with two methods (probing and steering) into a 3 × 2 task matrix, tested across a value concept dataset of ten dimensions.
4. An operational foundation for value-driven autonomous AI, proposed as a way to shape LLM reasoning and decision-making via learned value representations rather than purely behavioral training such as RLHF or Constitutional AI.

Main Findings

• Probing is value-specific, not generic: Across all three regimes (A-A, A-C, and C-C), diagonal cells (matched probe and corpus) show significantly higher probing scores than off-diagonal cells (mismatched pairs), indicating that trained probes capture value identity rather than generic semantic features.
• Cross-level transfer occurs: Probes trained solely on abstract value descriptions reliably detect the same values in concrete event narratives and in decision reasoning, demonstrating that value representations persist from pure interpretation through event grounding to decision rationale.
• The pattern holds across models: All six tested LLMs show significantly positive diagonal-minus-off-diagonal activation differences in all three regimes; full cross-validation matrices are reported in the appendix.
• Middle layers encode values best: Probing correlation rises in shallow layers, peaks in middle layers, and declines or fluctuates in deeper layers, suggesting intermediate layers are most effective at encoding abstract value representations.
• Steering works for grounding and decisions: For A-C and C-C, decreasing steering strength α suppresses the value representation — fewer corpora are judged value-relevant and mean probability falls; increasing α amplifies it, shifting the distribution upward. At neutral steering (α = 0), mean value-relevance judgment is around ∼50%.
• Steering fails for abstract interpretation: In A-A, mean judgment probability remains ∼50% regardless of steering strength. The authors attribute this to highly polarized abstract concept corpora (passages judged as either strongly relevant or strongly irrelevant, with few intermediate cases), and interpret the null result as evidence that encoded abstract values act as stable representational anchors rather than malleable activations — or alternatively that they lie in "safe regions" of the representation space.
• External validation supports the probes: Three advanced LLMs (GPT-5.2, Gemini-3-Pro, Claude-Sonnet-4.5) rated value relevance of the A-C corpora at the corpus level, and these ratings correlated highly with probe mean scores. The specific correlation values are not reported in the truncated text.

Methodology in Plain English

The authors built a dataset of ten value dimensions (patriotism, equality, integrity, cooperation, individualism, discipline, curiosity, courage, satiety, rest). Using GPT-4o, they generated 800 sentences per value — 400 strictly relevant and 400 irrelevant — then asked GPT-4o to elaborate its understanding of each sentence within 80 words, producing 800 interpretations that serve as abstract-level value descriptions. GPT-4o then scored every token in these texts on a 7-point scale from completely irrelevant (0) to highly relevant (6). The data was split 90% for probe training (720 sentences per value) and 10% for testing.

For probing, they extracted activations from the MLP output of each transformer layer and trained a linear probe per value per layer, using a ReLU output and minimizing mean squared error plus an L1 penalty on the probe weights. The probe is essentially a direction in activation space that fires more when the relevant value is present. To check specificity, they built matrices where each probe is applied to every value corpus: a valid probe should score higher on its matched corpus than on mismatched ones. The best-performing layer per value, chosen by Pearson correlation on validation data, was used as the "diagnostic probe" in all later experiments.

For steering, they reused the same probe direction as a lever: they added a scaled version of the probe vector (α × k_p × w_p) to the activations, where k_p normalizes the direction, and then measured how the model's probability judgments over options shifted. Passive probing establishes existence; active steering establishes causal efficacy. Experiments ran on two NVIDIA RTX 4090 GPUs.

Why This Matters

• Impact on research: The work reframes alignment as a question of internal representation rather than only output behavior, offering a mechanistic complement to behavioral methods like RLHF and Constitutional AI. The dissociation between A-A and A-C/C-C — where steering slides concrete judgments but not abstract interpretations — suggests abstract values may be stable anchors in representation space, which is a specific, testable claim about how values are encoded.

• Value-aligned autonomous agents: Agents operating in open-ended environments could be steered along value directions to regulate decisions rather than retrained on exhaustive behavioral data.

• Transparency and auditing: Probing provides a read-out of which values are active in a model's internal state, which could support auditing tools for deployed systems.

• Targeted behavioral control: Steering on grounding and decision tasks (A-C, C-C) offers a mechanism for shifting model choices in concrete scenarios without changing the model weights.

• Alignment evaluation: The cross-level probe methodology could be reused to test whether fine-tuning or prompting actually internalizes principles, not just mimics them.

• Industry relevance: The framework targets six widely available open-source models (Qwen3-4B, Qwen3-8B, Llama3-3B, Llama3-8B, Mistral-7B, Gemma2-9B) on consumer-grade hardware (two RTX 4090 GPUs), making internal-representation control broadly applicable to teams building or deploying open-weight models.

Future Directions

• Overcoming the single-layer probe limitation: The authors note that probe accuracy is inherently limited by the distributed nature of signals, and that averaging large amounts of diffuse signal is necessary even though the network remains sensitive to minor fluctuations.
• Mapping the effective domain of steering: Preliminary observations indicate steering fails when α is excessively high, but the authors state this warrants more comprehensive analysis.
• Explaining why abstract representations resist steering: The A-A null result is interpreted two ways — stable anchors versus "safe regions" in semantic space — and distinguishing these hypotheses is left open.
• Extending beyond the ten value dimensions and six models tested, and clarifying how the observed middle-layer encoding peak relates to multi-layer information processing.

Target Audience

Researchers and engineers working on LLM interpretability, AI alignment, and value alignment; practitioners who want mechanistic tools for steering open-source models; and anyone interested in whether abstract concepts such as human values are internally grounded in language models rather than merely manipulated as surface patterns.

Authors’ abstract

Do large language models (LLMs) genuinely understand abstract concepts, or merely manipulate them as statistical patterns? We introduce an abstraction-grounding framework that decomposes conceptual understanding into three capacities: interpretation of abstract concepts (Abstract-Abstract, A-A), grounding of abstractions in concrete events (Abstract-Concrete, A-C), and application of abstract principles to regulate concrete decisions (Concrete-Concrete, C-C). Using human values as a testbed - given their semantic richness and centrality to alignment - we employ probing (detecting value traces in internal activations) and steering (modifying representations to shift behavior). Across six open-source LLMs and ten value dimensions, probing shows that diagnostic probes trained solely on abstract value descriptions reliably detect the same values in concrete event narratives and decision reasoning, demonstrating cross-level transfer. Steering reveals an asymmetry: intervening on value representations causally shifts concrete judgments and decisions (A-C, C-C), yet leaves abstract interpretations unchanged (A-A), suggesting that encoded abstract values function as stable anchors rather than malleable activations. These findings indicate LLMs maintain structured value representations that bridge abstraction and action, providing a mechanistic and operational foundation for building value-driven autonomous AI systems with more transparent, generalizable alignment and control.